Deep neural networks have achieved remarkable success in a wide variety of
natural image and medical image computing tasks. However, these achievements
indispensably rely on accurately annotated training data. If encountering some
noisy-labeled images, the network training procedure would suffer from
difficulties, leading to a sub-optimal classifier. This problem is even more
severe in the medical image analysis field, as the annotation quality of
medical images heavily relies on the expertise and experience of annotators. In
this paper, we propose a novel collaborative training paradigm with global and
local representation learning for robust medical image classification from
noisy-labeled data to combat the lack of high quality annotated medical data.
Specifically, we employ the self-ensemble model with a noisy label filter to
efficiently select the clean and noisy samples. Then, the clean samples are
trained by a collaborative training strategy to eliminate the disturbance from
imperfect labeled samples. Notably, we further design a novel global and local
representation learning scheme to implicitly regularize the networks to utilize
noisy samples in a self-supervised manner. We evaluated our proposed robust
learning strategy on four public medical image classification datasets with
three types of label noise,ie,random noise, computer-generated label noise, and
inter-observer variability noise. Our method outperforms other learning from
noisy label methods and we also conducted extensive experiments to analyze each
component of our method.
This problem is even more severe in the medical image analysis field, as the annotation quality of medical images heavily relies on the expertise and experience of annotators.
In this paper, we propose a novel collaborative training paradigm with global and local representation learning for robust medical image classification from noisy-labeled data to combat the lack of high quality annotated medical data.
Notably, we further design a novel global and local representation learning scheme to implicitly regularize the networks to utilize noisy samples in a selfsupervised manner.
We evaluated our proposed robust learning strategy on four public medical image classification datasets with three types of label noise, i.e., random noise, computer-generated label noise, and inter-observer variability noise.
Our method outperforms other learning from noisy label methods and we also conducted extensive experiments to analyze each component of our method.
提案手法は,ノイズラベル法で他の学習よりも優れており,また,各成分の分析実験も行った。
0.74
Index Terms—Noisy label, collaborative training, representa-
指標項 -ノイズラベル、協調訓練、表現-
0.61
tion learning, self-supervision
自尊心学習, 自己スーパービジョン
0.55
I. INTRODUCTION
I. イントロダクション
0.64
Robust medical image analysis has been an increasingly important topic, as accurate and robust algorithms are desired to increase clinical workflow efficiency and reliably support medical-decision-mak ings.
With the advent of deep learning, encouraging human-level performance has been achieved in a spectrum of challenging medical image diagnosis applications including skin cancer [1], lung cancer [2], retinal diseases [3], histology image analysis [4], etc..
The success of these applications relies on highly discriminative representations learned by convolutional neural networks (CNNs) from a large amount of carefully labeled data with the aid of domain experts.
Despite the remarkable success, it has been frequently seen that the performance of CNNs is easily affected due to the bias of the training set in complex real-world situations.
For C. Xue, P. Chen, Q. Dou and P.A. Heng are with the Department of Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong, China (e-mail: xchengjlu@gmail.com) .
のために C. Xue, P. Chen, Q. Dou, P.A. Hengは、香港の中国大学、香港、中国(eメール:xchengjlu@gmail.com )のコンピュータ科学・工学部に所属している。
0.70
L. Yu is with the Department of Statistics and Actuarial Science, The University of Hong Kong, Hong Kong, China (email:lqyu@hku.hk).
L. Yuは香港大学 (香港) の統計・アクチュエーター科学部に所属している(email:lqyu@hku.hk)。
0.68
(Corresponding author: Lequan Yu.) Fig. 1: Typical examples of ambiguous malignant and benign images from histology lymph node data (Top) and skin lesion data (Bottom), which are easily to be incorrectly labeled in the practical annotation process.
instance, such bias could come from statistical distribution mismatch across different clinical sites, imbalance of the number of samples across different disease categories, and the variations of disease visual patterns across patient age or gender populations.
Notably, besides these data biases in the aspect of the image itself, the bias with annotation quality commonly exists, and is harmful to the supervised-learningb ased diagnostic solutions.
One typical issue is the problem of noisy labels associated with the collected data, especially for the ambiguous medical images which may confuse clinical experts.
For example, as illustrated in Fig 1, the malignant histology lymph node images present quite similar colors and structures to benign ones, which shows that noisy annotations are inevitable in medical image analysis in practice.
The straightforward solution is manually reducing the presence of incorrect labels.
簡単な解決策は、間違ったラベルの存在を手動で減らすことだ。
0.52
However, this procedure is expensive, timeconsuming and impractical.
しかし、この手順は高価で時間がかかり、実用的ではない。
0.57
The naive training paradigm with noisy samples is reported to be inadequate, since the CNNs are parameterized with a large model capacity and can eventually memorize all training sample pairs including wrongly-labeled ones [5], [6].
Although handling noisy labeled medical data is a crucial task in the era of deep learning, merely preliminary studies have been conducted in the medical image analysis field.
training strategy by explicitly modeling the label noise as a part of network architecture, and presented the application of binary classification for breast mammograms.
For example, Goldberger et al [12] estimated the transition matrix of noisy labels by adding an extra softmax operation on top of the original softmax layer.
These methods rely heavily on estimating the noisy class posterior.
これらの手法はノイズクラス後部の推定に大きく依存している。
0.57
However, the estimation error for noisy class posterior could be largely due to the randomness of label noise and the complicated data biases, which would lead the transition matrix to be poorly estimated.
To be free of estimating noisy label transition matrix, a promising direction is to train deep learning models with clean samples, which can be selected according to the noisy classifier’s prediction [14].
Intuitively, if the noisy samples could be automatically filtered out from the training database during learning, the model robustness on noisy labeled data is gained.
For instance, Jiang et al [15] proposed the MentorNet to train an extra network with a clean validation set to select clean instances from the entire training database.
In another aspect, Han et al [16] proposed co-teaching, which simultaneously trains two networks in a symmetric way to select small loss samples for training.
別の側面として、Han et al [16] は2つのネットワークを対称的にトレーニングし、トレーニング用の小さな損失サンプルを選択することを提案した。
0.68
However, in the co-teaching framework, only part of the training data can contribute to the learning process in each epoch, which leads to unwanted waste of the valuable data samples with “noisy” annotations.
In practice, medical images usually contain hard samples that also have large training loss due to different image quality/imbalanced class distribution.
We propose a novel co-training framework with global and local representation learning for effectively tackling the sample selection bias and adopting all of the training data without wasting it.
We simultaneously train two student-teacher networks and each student-teacher network learns from each other to alleviate the bias caused by the noisy label.
In particular, in each training cycle, the two student-teacher networks are trained independently for one epoch, followed by a noisy label filter (NLF) to automatically
Then the divided samples are crossly fed into the peer student-teacher network as input training data.
そして、入力訓練データとして、分割サンプルを相互にピア学生教師ネットワークに供給する。
0.74
During training, the student-teacher network directly utilizes the clean samples in a fully supervised learning manner such that those noisy labels make less influence on model optimization.
This scheme includes a local contrastive loss to regularize the sample local clustering by pulling the similar samples together and pushing different samples far away; and a global relationship consistency loss, applied on all the samples, to regularize the sample structure relationship consistency.
In the testing phase, the final prediction result is acquired by averaging the prediction of the two teacher models.
試験段階では、2つの教師モデルの予測平均化により最終予測結果を取得する。
0.80
Our main contributions can be summarized as follows: • We propose a novel self-supervised loss with global and local regularization to preserve the sample utilization efficiency and alleviate the overfitting of noisy labels.
• We develop a new self-ensemble co-training framework for robust medical image classification.
•ロバストな医用画像分類のための新しい自己センブルコトレーニングフレームワークを開発した。
0.82
We employ a student-teacher network instead of a single network to more robustly filter the clean samples and thereby eliminate the disturbance from noisy samples.
• The presented noisy label learning strategy has been extensively validated on three typical medical image noisy labels with public datasets, i.e., random noise, inter-rater variability, and computer generated noise.
The proposed method consistently improves the prediction accuracy across different datasets and outperforms other learning from noisy labeled data methods.
Existing literature on training with noisy labels focuses primarily on two directions.
既存のノイズラベルによるトレーニングに関する文献は主に2つの方向に焦点を当てている。
0.53
(1) estimate the transition matrix approaches ( [17]–[20]).
1) 遷移行列のアプローチ([17]–[20])を推定する。
0.59
Sukhbaatar et al.
Sukhbaatarなど。
0.53
[21] proposed a noise transformation to estimate the transition matrix, which needed to be periodically updated and is non-trivial to learn.
[21]は,周期的に更新される必要があり,学習が簡単でない遷移行列を推定する雑音変換を提案した。
0.81
Patrini et al [13] designed a two-stage framework to learn on the noisy labeled training data, where in the first stage the transition matrix was estimated, and a correction loss was employed in the second stage.
Patrini et al [13] はノイズラベル付きトレーニングデータから学習するための2段階のフレームワークを設計し,第1段階で遷移行列が推定され,第2段階で補正損失が採用された。
0.85
Similarly, Goldberger et al [12] proposed to estimate the latent label transition matrix using deep learning method in an end-toend manner by adding an extra layer.
The accuracy of the classifier trained on noisy labeled data can be improved by such accurate estimation.
ノイズラベルデータに基づいて学習した分類器の精度は, 高精度な推定により向上する。
0.72
However, the label transition matrix is hard to be estimated in reality due to complicated data biases.
しかし、ラベル遷移行列は複雑なデータバイアスのため、現実に推定することは困難である。
0.75
Moreover, the label noises in the medical image are usually instance dependent noise, which makes the estimation of transition matrix even challenging.
Han et al [16] presented a co-teaching learning paradigm by simultaneously training two models and removing the potential noisy samples in each mini-batch according to the training loss of each input data.
Han et al [16]は、2つのモデルを同時に訓練し、各入力データのトレーニング損失に応じて、各ミニバッチ内の潜在的なノイズサンプルを除去することで、共学学習パラダイムを提示した。
0.64
Besides, some recent works also adopted a semi-supervised learning strategy to address the noisy label problem ( [11], [24], [26]– [28]) by generating pseudo labels or adding regularization.
Our method uses a self-supervised strategy to effectively utilize the samples by specifically designed inter-patient relationship loss and local contrastive loss.
B. Noisy Label Learning for Medical Image Analysis
B. 医用画像解析のための雑音ラベル学習
0.90
The annotation quality of medical
医学のアノテーションの質
0.70
images is prone to experience, which requires years of professional training and domain knowledge.
イメージは経験しがちです 長年の専門的な訓練と ドメイン知識が必要です
0.75
There exist many ambiguous images that will confuse the clinical experts, and therefore, results in misdiagnosis/wrong annotation, and disagreements.
臨床専門家を混乱させる曖昧な画像が多数存在し、誤診・怒りの注釈や意見の相違が生じている。
0.64
For some large-scale datasets, the annotation is also heavily relied on automatically extracting labeling from radiological reports by natural language processing tools, which inevitably results in a certain level of label noise.
Manually reducing those incorrect annotations requires agreements between experts and is timeconsuming.
これらの誤ったアノテーションを手動で削減するには、専門家間の合意が必要であり、時間がかかります。
0.38
Some studies have addressed the existence of noisy labeled medical images during the network training process.
ネットワークトレーニング過程におけるノイズラベル付き医療画像の存在に対処する研究もある。
0.77
Dgani et al [7] proposed to estimate the transition matrix of the noisy label during the training process by optimizing an extra softmax layer, it was heavily dependent on the noise label transition matrix assumption.
Dgani et al [7] は, 付加ソフトマックス層を最適化することにより, ノイズラベル遷移行列の仮定に大きく依存し, トレーニング過程におけるノイズラベルの遷移行列を推定することを提案した。
0.77
Shu et al [10] proposed to weaken the influence of noisy annotation in segmentation tasks by utilizing potential visual guidance during the learning process.
To suppress the influence of noisy labels, some methods proposed a sample selection strategy to train deep learning models on selected data or assign appropriate weight to the training data.
Xue et al [8] proposed a robust learning framework for noisy labeled medical image classification, where the severe hard samples and noisy labels in medical images are both considered by a sample selection module and a sample re-weighting strategy.
Xue et al [8] は, 医用画像分類のための頑健な学習フレームワークを提案し, 医用画像中の強硬なサンプルとノイズのあるラベルを, サンプル選択モジュールとサンプル再重み付け戦略により検討した。
0.85
Min et al [29] presented a semi-supervised biomedical segmentation network with noisy labels, where the noisy label issue was tackled by weakening the noisy gradients in multiple layers.
Le et al [30] utilized a small set of clean training samples and assigned weights to training samples to deal with sample noise.
le et al [30]はクリーンなトレーニングサンプルのセットを使用し、サンプルノイズに対処するためにトレーニングサンプルにウェイトを割り当てた。 訳抜け防止モード: Le et al [30 ] はクリーントレーニングサンプルの小さなセットを利用した サンプルノイズに対処するため トレーニングサンプルに重量を割り当てました
0.86
However, the previously proposed sample selection strategy has not considered the selection bias caused by the erroneous guiding of noisy samples, and only part of the valuable training data is used.
(2) How to utilize the detected noisy labeled data.
2)検出されたノイズラベルデータの利用方法。
0.76
Some studies discarded samples with a high probability of being incorrectly labeled or maintain these samples in a semi-supervised manner by refining those labels ( [25], [26]).
Premaladha et al [35] adopted deep learning and hybrid AdaBoost Support Vector Machine (SVM) algorithms to classify melanoma skin lesions.
Premaladha et al [35]は、メラノーマ皮膚病変を分類するために、ディープラーニングとハイブリッドAdaBoost Support Vector Machine (SVM)アルゴリズムを採用した。 訳抜け防止モード: Premaladha et al [35 ] はディープラーニングとハイブリッド AdaBoost Support Vector Machine (SVM) アルゴリズムを採用した メラノーマの皮膚病変を分類します
0.82
This system was tested on 992 images and obtained a high classification accuracy (93%).
このシステムは992枚の画像でテストされ、高い分類精度(93%)を得た。
0.77
Kawahara et al [34] employed a fully convolutional network to extract multi-scale features for melanoma recognition, the author reported accuracy of 85.8% on a five-class classification problem.
Cruz et al [41] presented a classification approach for detecting the presence and extent of invasive breast cancer on the whole slide digitized pathology images using a CNN classifier.
Cruz et al [41] は,CNN分類器を用いて,スライド画像全体に対する浸潤性乳癌の存在と範囲を検出するための分類手法を示した。
0.85
A. Problem Setting III.
A.問題設定 III。
0.75
METHODS In supervised learning, we consider finding a mapping function f : X → Y, where X is image space, Y is label space, and f describes the complex relationship between them.
方法 教師あり学習では、x が画像空間、y がラベル空間、f がそれらの間の複素関係を記述する写像関数 f : x → y を見つけることを考える。
0.48
To learn the mapping function f (·), a loss objective is usually defined to penalize the observed differences between the model prediction f (x) and the ground truth target y for a training
写像関数 f(·) を学習するために、通常、損失目標がモデル予測 f(x) と基底真理目標 y との観測された相違をペナルティ化するために定義される。
0.79
英語(論文から抽出)
日本語訳
スコア
4 Fig. 2: An overview of the proposed co-training with global and local representation learning framework:
4 図2:グローバルおよびローカルな表現学習フレームワークによる協調学習の提案の概要
0.58
(a) is the co-training scheme with two independent student-teacher networks and a noisy label detection procedure.
(a)2つの独立した学生・教師ネットワークと雑音のラベル検出手順による共同学習方式である。
0.65
NLF is the noise label filter.
NLFはノイズラベルフィルタである。
0.86
(b) represents the detailed training procedure of the teacher-student network.
(b)教師学生ネットワークの詳細な訓練手順を表す。
0.78
The network directly employs the clean samples with supervised cross-entropy loss.
ネットワークは、監視されたクロスエントロピー損失を伴うクリーンサンプルを直接使用する。
0.58
The local contrastive loss is applied on noisy samples and the global relation loss is applied on both clean and noisy samples.
The real ground truth label y for the sample is unknown due to various annotation limitations, misdiagnosis, or disagreements.
サンプルの実際の基底真理ラベル y は、様々な注釈制限、誤診断、不一致のため不明である。
0.68
During training, we only have ˆy assigned to each sample, which means that for a noisy labeled training dataset, there are input images x, the noisy labels ˆy, and the latent clean labels y.
つまり、ノイズラベル付きトレーニングデータセットには、入力画像x、ノイズラベルs_y、潜在クリーンラベルyがある。 訳抜け防止モード: トレーニング中は、各サンプルに割り当てられた'y'しかありません。 つまり ノイズラベル付きトレーニングデータセットには 入力画像 x, ノイズラベルは y で、遅延クリーンラベルは y です。
0.65
In our approach, we train a network based on {x, ˆy} with underlying {x, y} unavailable.
It has been reported that naive training with the noisy labels ˆy will result in performance degradation for predictions on test set ( [5], [6]).
ノイズラベル sy によるナイーブなトレーニングは、テストセット ([5], [6]) における予測のパフォーマンス低下をもたらすと報告されている。 訳抜け防止モード: ノイズラベル sy によるナイーブなトレーニングは,テストセット ([5]) における予測性能の低下をもたらすと報告されている。 [ 6 ] ) .
0.76
With our proposed robust learning strategy, we aim to optimize the CNN classifier with {x, ˆy} while achieving comparable results with a model trained on {x, y}.
Our framework maintains two independent teacher-student networks trained simultaneously, denoted as student-teacher A and student-teacher B. These two networks have different weight parameters due to the different initialization, different image augmentation, different training data division, and different mini-batch sequences so that they can filter out the noisy labels independently.
これら2つのネットワークは,初期化,画像拡張,トレーニングデータ分割,およびミニバッチシーケンスの違いにより,それぞれが独立してノイズラベルをフィルタリングできるように,異なる重みパラメータを持つ。 訳抜け防止モード: 我々のフレームワークは2つの独立した教師-学生ネットワークを同時に訓練している。 学生 - 教師 - A と 生徒 - 教師 - B で表される。 異なるイメージ拡張、異なるトレーニングデータ分割、異なるミニ - バッチシーケンスのために ノイズラベルは別々にフィルタリングできます
0.77
Fig 2(a) shows the procedure of one training cycle, after train the
(b)Student-TeacherNe tworkwithglobalandlo calrepresentationlea rningPredicted Loss (A)Predicted Loss (B)Co-training withdivided dataClassifierClassi fierConsistencyLossC leanNoisyStudentEnco der
b)Student-TeacherNet workwithglobaland Localrepresentationl earning Predicted Loss (A)Predicted Loss (B)Co-training withdivided dataClassifierClassi fierConsistencyLossC leanNoisyStudentEnco der
0.15
英語(論文から抽出)
日本語訳
スコア
student-teacher network for one epoch, all the training data go through a noisy label detection process, where the predicted loss (A) and predicted loss (B) of all the samples on studentteacher A and student-teacher B are calculated.
Different from Co-teaching [16] that selected small loss samples and proceed crossed training in each mini-batch, we perform crossed training after each epoch with the divided training samples.
To tackle the challenge of noisy label detection, after each epoch we predict the cross entropy loss of all the training data by two teacher networks respectively and input the cross entropy loss into a noisy label filter (NLF) separately as shown in Fig 2 (a).
The NLF adopts a two-component Gaussian Mixture Model to fit the max-normalized loss of the training data using the Expectation-Maximiza tion algorithm in which the clean and noisy components are divided in an unsupervised manner.
During training, we employ a linear schedule t, that t gradually decreased from 0.9 to 0.5 in the first 10 epochs, and maintain 0.5 in the following epochs.
訓練中は線形スケジュールtを用い,最初の10エポックでは0.9から0.5に徐々に減少し,次のエポックでは0.5を維持した。 訳抜け防止モード: トレーニング中, 線形スケジュール t を用い, t は 0.9 から 0.5 へと徐々に減少していった。 以下のエポックで0.5を維持します
0.81
The strategy can select the most confident samples at the initial stage, then gradually reduce the confidence requirement as the model is becoming more robust.
The initial threshold is 0.9 for the histology lymph node images since it is a balanced dataset, while for the other dataset, we set the initial threshold to 0.8 to avoid neglect of less represented class.
Fig 2(b) shows the training scheme of the student-teacher network, which adopts the augmented clean and augmented noisy dataset as input for the student encoder and teacher encoder.
The whole network is trained by four losses that will be introduced in the next section.
ネットワーク全体が次のセクションで導入される4つの損失によってトレーニングされている。
0.77
The studentteacher network is proposed for semi-supervised learning, while we adopt it in the noisy label learning scenario.
セミ教師付き学習には学生教師ネットワークが提案され,ノイズの多いラベル学習シナリオでは採用される。
0.68
As we only update the student network with the selected small loss samples and update the weights of the teacher model by the exponential moving average (EMA) of student network weights.
Since we observe that the deep learning model always first learns the easy instances and adapts to hard noisy instances, as the separable of clean and noisy data gradually decreased as shown in Fig 3.
After the noisy label detection procedure, instead of discarding the noisy labeled sample, we propose to utilize the noisy labeled samples by the self-supervised learning strategy, which can mitigate the requirement for labeled data by providing a means of leveraging unlabeled data.
The regularization is designed in two perspectives: the global domain, where we align the persample relation matrix of each batch to preserve the general knowledge of the inter-sample relationships; and the local domain, where we add a contrastive loss to the noisy labeled samples [43] to strength the feature representation learning.
We calculate an inter-sample relation matrix that can capture the relationships between each sample within a mini-batch.
ミニバッチ内で各サンプル間の関係をキャプチャできるサンプル間関係行列を計算する。
0.75
Specifically, for each mini-batch with B samples, we extract the feature before the f c layer of the backbone network, the shape of the feature map in ResNet34 is F ∈ RB×512.
具体的には、Bサンプルを持つ各ミニバッチに対して、バックボーンネットワークのfc層の前に特徴を抽出し、ResNet34における特徴写像の形状は F ∈ RB×512 である。
0.75
The relation matrix is computed as:
関係行列は次のように計算される。
0.59
G = F · F T
G = F · F T
0.42
(2) where Gij is the inner product between the vectorized activation map F
(2) ここで Gij はベクトル化活性化写像 F の間の内積である
0.51
(i) and F (j), whose intuitive meaning is the similarity between the ith sample and jth sample within the input mini-batch.
(i)とF (j)は入力されたミニバッチ内のithサンプルとjthサンプルの類似性である。
0.56
The final sample relation matrix M is obtained by conducting the L2 normalization for each row Gi of G. We then propose a global inter-sample relationship loss Lglobal to align the inter-sample relationship of samples between the teacher model and student model by minimizing their symmetrized Kullback–Leibler (KL) divergence
G の各行 Gi に対して L2 正規化を実施して最終サンプル関係行列 M を得る。次に,その対称性を最小化した Kullback-Leibler (KL) 分散を最小化し,教師モデルと学生モデルの間のサンプル間関係を整合させる大域的なサンプル間関係損失 Lglobal を提案する。
0.75
(DKL(M s||M
(DKL(M s||M)
0.35
t) + DKL(M t||M
t) + DKL(M t||M)
0.78
s)), (3) Lglobal =
s)。 (3) Lglobal =
0.44
where DKL(p||q) = (cid:80)
ここで dkl(p||q) = (cid:80)
0.52
1 2 i pilog pi qi
1 2 I pilog pi qi
0.41
, M s and M t represent the sample relation matrix M of student network and teacher network respectively.
M s と M t はそれぞれ学生ネットワークのサンプル関係行列 M と教師ネットワークを表す。
0.77
This global inter-sample relationship loss is applied to all the samples.
このグローバルなサンプル間の関係損失は、すべてのサンプルに適用される。
0.69
Local contrastive loss.
ローカルなコントラスト損失。
0.70
In addition to promoting the alignment of feature relationships with Lglobal, we further encourage robust feature representation by adding a local self-supervised contrastive loss following [43], denoted as Llocal.
The projection layer includes one hidden layer and ReLU.
投影層は、1つの隠れ層とreluを含む。
0.68
The local contrastive loss is calculated for each sample in the mini-batch.
ミニバッチの各サンプルに対して局所的なコントラスト損失を算出する。
0.70
We augment each image to two images, so that the difference between positive pairs (two augmented images from same source image) can be minimized while the difference from negative pairs (augmented images from different source) is maximized.
(4) where zi and zj denote the projection of features, τ is a scalar temperature parameter.
(4) ziとzjが特徴の投影を表す場合、τはスカラー温度パラメータである。
0.58
We analyze the influence of τ in the i zj/||zi||||zj|| denote the experiment section.
実験区間を表す i zj/||zi|||zj|| における τ の影響を分析する。
0.58
sim(zi, zj) = zT cosine similarity between zi and zj.
sim(zi, zj) = ztコサインの類似性 ziとzjの間の類似性。
0.73
D. Overall Loss Functions Given the batches in each iteration, the total loss consists of cross-entropy loss LCE between the observed labels and model predictions for the clean dataset and the unsupervised loss, which includes the original consistency loss Lcon of mean-teacher for both clean and noisy labeled dataset, the local contrastive loss Llocal for noisy labeled data, and global inter-sample relationship alignment loss Lglobal for both clean and noisy labeled data.
d. 総損失関数 各イテレーションのバッチが与えられると、総損失は、観測されたラベル間のクロスエントロピー損失lceと、クリーンデータセットとノイズラベル付きデータセットのための平均教師の平均一貫性損失lconと、ノイズラベル付きデータのためのローカルコントラスト損失llocalと、クリーンデータとノイズラベル付きデータの両方に対するグローバルグループ間アライメント損失lglobalとからなる。
0.80
Specifically, the loss function for clean data is the cross entropy loss
具体的には、クリーンデータの損失関数はクロスエントロピー損失である
0.83
LCE = − 1 Nclean
LCE = − 1 Nclean
0.43
(cid:88) (cid:88)
(cid:88) (cid:88)
0.39
i∈Nclean j∈C
I-Nclean j・C
0.33
yj i log f j
yj i log f j
0.42
s (xi, θ),
s (xi, θ) である。
0.78
(5) (cid:88)
(5) (cid:88)
0.41
i∈N where Nclean is the number of clean samples and C is the number of class.
いはーん Ncleanはクリーンなサンプルの数で、Cはクラスの数です。
0.56
Also, the original consistency loss is adopted
また もともとの一貫性の損失は
0.66
Lcon = 1 N ||fs(xi, θs) − ft(x(cid:48)
Lcon = 1N ||fs(xi, θs) − ft(x(cid:48)
0.42
i, θt))||2 2,
i, θt))||2 2 である。
0.69
(6) where x and x(cid:48) represent two differently augmented input images, θs and θt are the weights of student model and teacher model, N is the total image number.
(6) x と x(cid:48) は2つの異なる拡張された入力画像を表し、θs と θt は学生モデルと教師モデルの重みを表し、N は総画像数である。
0.59
Then the total objective function can be represented as
すると、全目的関数を表現できる。
0.71
Ltotal = LCE + λ(Lglobal + Llocal + Lcon),
Ltotal = LCE + λ(Lglobal + Llocal + Lcon)
0.39
(7) where λ is the weight of unsupervised loss.
(7) ここで λ は教師なし損失の重さです
0.53
Follow the common practice ( [45], [46]), we linearly ramp up λ from 0 to 10 over the first 10 epochs after warm-up procedure.
The exponential moving average decay (EMA) of the mean-teacher network was 0.99 according to [45].
平均教師ネットワークの指数移動平均減衰(EMA)は[45]により0.99であった。
0.86
The temperature of local loss τ was 0.5, and unsupervised loss weight λ was 10.
局所損失 τ の温度は0.5 であり,教師なし損失重量 λ は 10。
0.81
We adopted the ImageNet pre-trained ResNet34 with randomly initialized fully connected layer as the backbone for the skin lesion, lymph node, and gleason grading datasets, DenseNet121 as the backbone for NIH dataset.
IV. EXPERIMENTS To evaluate the effectiveness of our proposed self-ensemble co-training framework for noisy label learning, we have conducted extensive experiments on four challenging medical image diagnosis tasks:
The Kaggle histopathologic cancer detection dataset is a slightly modified version of the PatchCamelyon (PCam) [4], [47], which removed the duplicate images.
The second dataset is the Gleason 2019 dataset that contains 333 tissue microarray (TMA) images of prostate cancer, which are sampled from 231 radical prostatectomy patients.
We first evaluated our method on two synthetic noisy label datasets.
提案手法を2つの合成ノイズラベルデータセットで評価した。
0.64
To simulate the noisy label situations on our employed melanoma dataset and lymph node dataset, we randomly sampled γ ∈ {0.05, 0.1, 0.2, 0.4} percentage of images from each class and flip the labels of these images following the common setting.
The noisy label is defined as y(cid:48) i = yi with the probability of 1 − γ, and y(cid:48) i = yk, yk (cid:54)= yi with the probability of γ, where y(cid:48) i is the corrupted noisy label, yi is clean label.
雑音ラベルは 1 − γ の確率で y(cid:48) i = yi と定義され、y(cid:48) i = yk, yk (cid:54)= yi は γ の確率で定義される。 訳抜け防止モード: 雑音ラベルは 1 − γ の確率で y(cid:48 ) i = yi と定義される。 そして y(cid:48 ) i = yk, yk (cid:54) = yi の確率は γ である。 y(cid:48 ) iは破損したノイズラベルです Yiはクリーンなラベルだ
0.84
We adopted the accuracy as evaluation metric.
我々はその精度を評価基準として採用した。
0.51
Computer-generated noisy label.
コンピュータ生成ノイズラベル。
0.69
We trained our method on the NIH Chest X-ray dataset to simulate the reality and then further tested our trained model on a manually relabeled chest X-ray dataset to evaluate the effectiveness of our method.
Fig. 4: The Gleason grading from different pathologists for one sample histological image.
図4:Gleasonは、異なる病理学者から1つのサンプル組織像に分類される。
0.69
chance agreement and 1.0 means perfect agreement) [50], [51], as shown in Fig 4.
チャンス合意と1.0は完全合意を意味する) [50], [51] 図4に示すとおり、
0.74
Following [51], [52], we applied the Simultaneous Truth and Performance Level Estimation (STAPLE) algorithm [53], which is based on the expectationmaximizat ion method to estimate the ground-truth labels from the annotations of six radiologists.
ELR Ours A 78.52 76.75 79.80 80.66 79.58 80.37 81.76
我々の A 78.52 76.75 79.80 80.66 79.58 80.37 81.76
0.36
B 66.47 68.92 68.21 69.63 71.93 70.27 75.47
B 66.47 68.92 68.21 69.63 71.93 70.27 75.47
0.32
• ELR [28]: They leveraged semi-supervised learning techniques and regularization term to prevent memorization of false labels.
• ELR [28]: 半教師付き学習手法と正規化項を活用し, 偽ラベルの暗記を防止する。
0.74
We evaluated the classification performance of these methods and our method by calculating their average classification accuracy or area under curve (AUC).
我々は,これらの手法の分類性能を,平均分類精度または曲線下面積(AUC)を算出して評価した。
0.83
The results are listed in Table I, Table II, Table III, and Table IV for the four datasets with three noise types.
For fair comparisons, we obtained the classification results of the comparison methods by downloading their public implementations or reimplementing the method according to their paper.
For all the noise ratio settings, decoupling obtained less improvement over cross entropy loss, indicating that even with two network settings, the training strategy that updates the network according to disagreements may not work well in the medical image domain.
For the noisy label dataset (noise ratio > 0), Both Co-teaching and our method outperformed other methods by a large margin, indicating that the sample selection bias is efficiently alleviated by the two-network cotraining scheme.
Furthermore, our method outperformed Coteaching in all the noise settings, showing that by adding the noisy label detection and two unsupervised loss can learn more robust feature representation.
For the small noise ratio setting, the overfitting to noisy label is not as severe as the heavily corrupted dataset, but all the methods achieved better results than training with cross entropy, and our method has the best results.
Note that the skin lesion classification data is more challenging than the lymph node classification task, as the data is very imbalanced and contains more hard samples.
when there is no noise in the training data, Decoupling and Mentornet obtained lower accuracy than cross entropy loss, which is reasonable as Decoupling only utilized partial data for training and Mentornet possibly assigned smaller weight for hard samples.
Those filtered or down-weighted samples are crucial for discriminative feature learning.
フィルタリングまたはダウンウェイトのサンプルは、識別的特徴学習に不可欠である。
0.58
In 0.05 noise ratio setting, Co-teaching does not perform well in the mild case due to the low sample utilization rate and wrong identification of hard samples; while self-paced Mentornet did not perform well due to the sample selection bias by single network setting and the down-weighted hard samples.
F-correction obtained marginal improvement on accuracy when the noise ratio is small, but the performance dropped dramatically when the noise ratio reached 0.4, as the transition matrix estimation is not robust in this setting.
F補正はノイズ比が小さい場合の精度において限界改善を得たが、遷移行列推定が堅牢でないため、ノイズ比が0.4に達すると性能が劇的に低下した。 訳抜け防止モード: F - ノイズ比が小さい場合の精度の限界改善を得た補正。 しかし ノイズ比0.4で 劇的に低下し この設定では 遷移行列の推定は堅牢ではありません
0.79
Compared with our methods, the Co-teaching entirely removed the selected noisy label which would result in a worse classification performance due to the missing of hard samples.
Results demonstrate that our method can effectively work on the challenging imbalanced dataset, as the co-training scheme with noisy label filter can filter out the noisy labels and the two self-supervised losses maintain an effective utilization of hard samples (minority group).
Manual annotation of large-scale datasets is time-consuming and expensive, some public datasets are often extracted by NLP from radiological reports, which inevitably results in a certain level of label noise.
Wang et al [48] reported the text mining accuracy of NIH Chest X-ray dataset on 900 manual labeled images, which shows that the precision and recall for Nodule and Pneumonia is below 90%.
Wang et al [48]は、900個の手動ラベル付き画像上にNIH Chest X-rayデータセットのテキストマイニング精度を報告し、NoduleとPneumoniaの精度とリコールが90%未満であることを示した。
0.75
Recently, Majkowska et al [49] have relabeled a subset (1,962 images) of NIH testing data with at least three radiologists per image.
Table III shows the AUC values of our method and other comparison methods.
表IIIは,本手法と他の比較手法のAUC値を示す。
0.79
Some of the comparison methods performed worse than the baseline, either for Pneumothorax or Nodule/Mass, we attribute this to the small noise proportion and unbalanced sample distribution of the 14 diseases, which makes the transition matrix estimation and sample selection challenging.
As our method integrated the noisy sample selection method and self-supervised learning strategy, it still outperformed the baseline and other methods.
Overall, this experiment demonstrates the robustness of the proposed method to label noise in the real clinical setting.
本実験は, 実際の臨床環境におけるノイズをラベル付けする手法の堅牢性を示すものである。
0.75
Results on public Gleason grading datasets.
公開gleason gradingデータセットの結果。
0.76
To evaluate our method on noisy data, we consider the grading by one patholo-
雑音データに基づく手法を評価するために,1つのパノロによる階調の検討-
0.58
英語(論文から抽出)
日本語訳
スコア
9 Fig. 5: (a) The performance of our method under different unsupervised loss weights on the histopathologic dataset.
9 第5図 a) 病理組織学的データセットにおける異なる教師なし損失重み下での手法の性能について検討した。
0.49
; (b) The AUC value of selected clean training samples under different noise setting, compared between ours and one network setting without Lglobal and Llocal on the lymph node dataset.
; b) Lglobal と Llocal を伴わない1つのネットワーク設定と比較し, 異なる雑音条件下で選択したクリーントレーニングサンプルのAUC値について検討した。
0.58
; (c) The change of global inter-sample relationship during network training under different scenarios.
; c)異なるシナリオにおけるネットワークトレーニングにおけるグローバルサンプル間関係の変化。
0.66
TABLE V: Ablation study on skin lesion and lymph node (Accuracy, %)
TABLE VI: The influence of temperature scale in contrastive loss (Accuracy, %).
表 vi: 比較損失における温度スケールの影響 (正確性, %)。
0.77
τ Skin lesion Lymph node
τ 皮膚病変リンパ節
0.55
0.01 84.50 91.88
0.01 84.50 91.88
0.24
0.1 84.17 91.25
0.1 84.17 91.25
0.24
0.5 84.33 92.75
0.5 84.33 92.75
0.24
1 85.17 92.25
1 85.17 92.25
0.34
gist as the noisy label and the label estimated by STAPLE from six pathologists as the clean label.
gistはノイズラベルであり、stapleは6人の病理学者からクリーンラベルと推定されている。
0.59
We conducted two groups of studies, A: all the training data and testing data are clean, B: we use the labels from one of the six pathologists as training data (noisy) and the testing data is clean.
All the methods show improvements over the baseline except F-correction.
全ての手法はF補正を除いてベースラインよりも改善されている。
0.58
Our method shows the best results among all the comparison methods.
提案手法は,すべての比較手法の最良の結果を示す。
0.75
D. Ablation Analysis Efficiency of key components.
D. キーコンポーネントのアブレーション解析効率
0.81
To demonstrate the effectiveness of our proposed method, we conducted the ablation study using both the histopathologic cancer detection dataset and ISIC skin dataset.
The results for each setting is shown in Table V. Our baseline is the ResNet34 trained by cross entropy loss and the result is provided in the first row of Table V. We mainly analyzed the efficiency of four components: co-training with noisy label filter, self-ensemble, global inter-sample relationship alignment loss, and the local contrastive loss.
From the ablation study, we observed that each component plays its own role in a complementary way.
アブレーション研究から,各成分が相補的な役割を担っていることを観察した。
0.73
Specifically, the cotraining with a noisy label filter scheme improves the accuracy
具体的には、ノイズラベルフィルタ方式によるコトレーニングにより精度が向上する。
0.59
by around 1%-4.7% compared to the baseline.
基準値の約1%-4.7%の値です
0.71
It shows that the two networks cross updated by selected samples yield improvement over the baseline by reducing the domination of noisy labels and alleviating the sample selection bias.
Then we evaluated the efficiency of the self-ensemble setting.
そして, 自己センブル設定の効率評価を行った。
0.74
By using the self-ensemble model (mean teacher network), the accuracy improved 1.6% − 7.5% since the teacher network is a more robust moving average model of the student network.
Hence, it is effective to adopt the self-ensemble model to select samples.
したがって, 自己感覚モデルを用いてサンプルを選択することは効果的である。
0.71
Lastly, we evaluated the effectiveness of the two proposed self-supervised losses.
最後に,提案する2つの自己監督損失の有効性を評価した。
0.56
As shown in line 4 and line 5 of Table V, only using the global inter-sample relationship alignment loss lglobal and the local contrastive loss Llocal achieved 1.7%−14.5% improvement on accuracy, indicates the effectiveness of these two losses.
It indicates that the representation learning capability of the network has been improved by training the network with global loss and local loss on the detected noisy labeled data.
We analyzed how the temperature hyperparameter in the local contrastive loss influences the network performance.
局所コントラスト損失における温度ハイパーパラメータがネットワーク性能に与える影響を解析した。
0.81
We ranged t ∈ [0.05, 0.1, 0.5, 1] for both datasets under the 10% noise ratio setting.
10%雑音比設定の下で, t ∈ [0.05, 0.1, 0.5, 1] の範囲を調べた。
0.62
As shown in Table VI, the classification accuracy generally improves the performance over the baseline (83.01% for skin and 88.95% for lymph node) while not being very sensitive to the value of t.
(a) (b) Fig. 6: The variation of filtered labeled data number.
(a) (b) 第6図:フィルタ付きラベル付きデータ番号のバリエーション。
0.55
(20 % and 40% noisy lymph node data)
(20%と40%のノイズ性リンパ節データ)
0.81
we ranged λ ∈ [0, 1, 10, 20] and observed the classification results on the histopathologic dataset.
λ ∈ [0, 1, 10, 20] の範囲で組織学的データセットの分類結果を観察した。
0.77
In Fig 5 (a), the bar-plots presents the mean of classification accuracy under all noise settings.
図5では (a)バープロットは、すべてのノイズ設定における分類精度の平均を示す。
0.82
We observed that our method generally improves the classification performance when λ < 10, while a larger weight of the self-supervision will overwhelm the supervised one and lead to underfitting.
In Fig 5 (b), we also plot the AUC value of the filtered clean samples at the last epoch.
図5では (b) 前回期における濾過洗浄試料のAUC値もプロットした。
0.72
Our method can consistently select high-quality samples with a higher AUC value compared to using one network to select samples, demonstrating that the selection procedure in our method is more robust.
The difference (variation) of filtered clean label sample numbers from two networks is presented in Fig 6.
フィルタされた2つのネットワークからのクリーンラベルサンプル数の差(変量)を図6に示す。
0.85
The two networks consistently selected a different number of samples, which shows the two networks have different learning abilities for noisy sample detection.
The results show that the loss of inter-sample relationships would not naturally converge during the network training when there is no explicit guidance.
The global alignment loss can enhance the feature robustness by aligning the relationship.
グローバルアライメント損失は、関係を整列することで特徴の堅牢性を高めることができる。
0.68
V. DISCUSSION Deep neural networks usually require large-scale annotated training data, while the annotations of medical images usually are prone to experience and disagreement may appear between doctors.
For diagnosis tasks in the medical image analysis field, the collection of large scale accurate annotated data is challenging due to the privacy policy and high work loading.
Furthermore, the hard sample is easy to be treated as the noisy label in medical image analysis, either discarding or re-labeling these data will lead to worse performance.
In this work, through the selfensemble co-training scheme with a robust noisy label filter and two tailored self-supervision loss , a novel deep learning framework was proposed for the noisy labeled medical image dataset.
Inspired by the collaborative training scheme proposed by [16], [42], a co-training network with a noisy label filter has been utilized to more explicitly enhance the network robustness and greatly alleviate the selection bias, as shown in Table V. Moreover, we incorporate a novel self-supervised learning scheme towards the global relationship alignment of samples in each batch and local clustering of each sample, which could maximally utilize the detected noisy labeled dataset (e g , noisy label or hard samples).
Inspired by the collaborative training scheme proposed by [16], [42], a co-training network with a noisy label filter has been utilized to more explicitly enhance the network robustness and greatly alleviate the selection bias, as shown in Table V. Moreover, we incorporate a novel self-supervised learning scheme towards the global relationship alignment of samples in each batch and local clustering of each sample, which could maximally utilize the detected noisy labeled dataset (e g , noisy label or hard samples). 訳抜け防止モード: 16], [42] によって提案された協調訓練計画に触発された ノイズラベルフィルタを用いたco-training networkは,ネットワークのロバスト性を高めるために活用されている 表 vに示すように 選択バイアスを大幅に緩和します 各バッチにおけるサンプルのグローバルリレーションアライメントと各サンプルの局所クラスタリングに向けて,新しい自己教師あり学習方式を導入する。 検出されたノイズラベル付きデータセット(ノイズラベルやハードサンプルなど)を最大限に活用することができる。
0.82
The experimental results demonstrate that this strategy efficiently improves the utilization of training data and improves the model performance.
Fig 7 shows that our self-ensemble network with two regularization losses could detect the noisy label more efficiently by properly dividing the noisy and clean dataset.
To evaluate the robustness of our method, it was tested on three types of noisy labels, i.e.random noise, computer generated label, and inter-observer variability.
For the random 010203040Epochs02550 75100125150175200Var iation0.20010203040E pochs050100150200250 Variation0.400.00.51 .00102030400.00.51.0 02468100.00.51.00123 0.00.51.001020304050 0.00.51.00204060800. 00.51.001020304050Cl eanNoise
The results suggested that our method can handle random noise even with very imbalanced data distribution.
その結果,不均衡なデータ分布であってもランダムノイズを処理できることがわかった。
0.74
To better evaluate the proposed method in a real-world setting, we trained the model using the public NIH datasets with NLP mined label and tested it on a manually labeled Chest X-ray dataset to verify the effectiveness of our method in reality.
Our method consistently shows better results than the baseline and other comparison methods, showing that the proposed method can handle the computer generated label with mild noise.
The proposed method was also tested on the Gleason grading dataset with multi-annotators, which is labeled by six pathologists with different experiences.
The dataset exists great inter-observer disagreements, our experiment showed that when using STAPLE estimated label as clean testing label, different training labels showed significantly different test results.
Our method achieved the best results among all the comparison methods with significant improvements (p < 0.05, paired t-test), as shown in Table VII and Table VIII.
本手法は,Table VII と Table VIII に示すように,有意な改善 (p < 0.05, paired t-test) を施した全比較法で最高の結果を得た。
0.82
Overall, we show that the co-training strategy and selfsupervised regularization can work together to build a robust training scheme, which successfully selected the high quality data for training, as shown in Fig 5(b).
Compared with other state-of-the-art methods, our method achieves significant improvement, no matter on a mild corrupted dataset or a heavily corrupted dataset.
In this study, we didn’t consider the disagreements between different annotators as a parameter during training, as recruiting annotations from multiple radiologists/patholo gists is timeconsuming in clinical practice.
In the future, we would like to extend our method to sample dependent label noise, such as considering the grading of each doctor as a clue for the noisy labels.
VI. CONCLUSION In this paper, we present a global and local representation guided co-training strategy to address the challenging yet important noisy label issue for medical image analysis.
The proposed method does not rely on refining or relabeling the noisy labeled data but employs two self-supervised losses to promote the learning of robust representation features.
The proposed framework can be easily extended to multi-class classification tasks and used in general classification networks for improving model robustness.
We extensively evaluated our method on four challenging medical
4つの挑戦医療の方法を 広く評価しました
0.68
REFERENCES [1] A. Esteva, B. Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. Blau, and S. Thrun, “Dermatologist-level classification of skin cancer with deep neural networks,” Nature, vol.
参考 [1] A. Esteva, B. Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. Blau, S. Thrun, “Dermatologist-level classification of skin cancer with deep neural network”, Nature, vol。 訳抜け防止モード: 参考 [1]A. Esteva, B. Kuprel, R. A. Novoa J. Ko, S. M. Swetter, H. M. Blau S. Thrun, “皮膚科医 – 深層神経ネットワークを用いた皮膚がんのレベル分類”。 自然。
0.64
542, no. 7639, p. 115, 2017.
542年、no.7639、p.115、2017年。
0.79
[2] A. A. A. Setio, A. Traverso, T. De Bel, M. S. Berens, C. van den Bogaard, P. Cerello, H. Chen, Q. Dou, M. E. Fantacci, B. Geurts et al , “Validation, comparison, and combination of algorithms for automatic detection of pulmonary nodules in computed tomography images: the luna16 challenge,” Medical image analysis, vol.
A. A. A. Setio, A. Traverso, T. De Bel, M. S. Berens, C. van den Bogaard, P. Cerello, H. Chen, Q. Dou, M. E. Fantacci, B. Geurts et al , “Validation, comparison, combination of algorithm for Automatic detection of lung nodules in computerd tomography image: the luna16 Challenge”, Medical image analysis, vol. 訳抜け防止モード: A. A. A. Setio, A. Traverso, T. De Bel M.S. Berens, C. van den Bogaard, P. Cerello, H. Chen Q. Dou, M. E. Fantacci, B. Geurts et al。 「CT画像における肺結節の自動検出アルゴリズムの検証・比較・組み合わせ : Luna16 チャレンジ」 医用画像解析。
0.76
42, pp. 1–13, 2017.
42, pp. 1-13, 2017年。
0.73
[3] J. De Fauw, J. R. Ledsam, B. Romera-Paredes, S. Nikolov, N. Tomasev et al , “Clinically applicable deep learning for diagnosis and referral in retinal disease,” Nature medicine, vol.
J. De Fauw, J. R. Ledsam, B. Romera-Paredes, S. Nikolov, N. Tomasev et al , “網膜疾患の診断とレファレンスのための臨床応用ディープラーニング”, Nature Medicine, Vol.
0.84
24, no. 9, p. 1342, 2018.
24.9.p.1342、2018年。
0.53
[4] B. E. Bejnordi, M. Veta, P. J. Van Diest, B. Van Ginneken, N. Karssemeijer, G. Litjens, J. A. Van Der Laak, M. Hermsen, Q. F. Manson, M. Balkenhol et al , “Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer,” Jama, vol.
B.E. Bejnordi, M. Veta, P. J. Van Diest, B. Van Ginneken, N. Karssemeijer, G. Litjens, J. A. Van Der Laak, M. Hermsen, Q. F. Manson, M. Balkenhol et al , “乳癌患者のリンパ節転移を検出するためのディープラーニングアルゴリズムの診断的評価” Jama, vol.
0.87
318, no. 22, pp. 2199–2210, 2017.
318, no. 22 pp. 2199-2210, 2017 頁。
0.81
[5] C. Zhang, S. Bengio, M. Hardt, B. Recht, and O. Vinyals, “Understand-
C. Zhang, S. Bengio, M. Hardt, B. Recht, O. Vinyals, “Understand”
0.38
ing deep learning requires rethinking generalization,” in ICLR, 2017.
ing deep learningは、2017年のiclrで、“一般化を再考する必要がある”。
0.57
[6] P. Chen, B. Liao, G. Chen, and S. Zhang, “Understanding and utilizing deep neural networks trained with noisy labels,” arXiv preprint arXiv:1905.05040, 2019.
P. Chen, B. Liao, G. Chen, S. Zhang, “Understanding and utilizeing Deep Neural Network trained with noisy labels” arXiv preprint arXiv: 1905.05040, 2019”。 訳抜け防止モード: [6 ]P. Chen, B. Liao, G. Chen, とS.Zhangは言う。 ノイズラベルで訓練されたディープニューラルネットワークを理解して活用する”。 arXiv preprint arXiv: 1905.05040 , 2019
0.71
[7] Y. Dgani, H. Greenspan, and J. Goldberger, “Training a neural network based on unreliable human annotation of medical images,” in Biomedical Imaging (ISBI 2018), 2018 IEEE 15th International Symposium on.
Y. Dgani, H. Greenspan, J. Goldberger, “Training a neural network based on unreliable human annotations of medical images” in Biomedical Imaging (ISBI 2018), 2018 IEEE 15th International Symposium on.[7] Y. Dgani, H. Greenspan, J. Goldberger。 訳抜け防止モード: [7]Y. Dgani、H. Greenspan、J. Goldberger 医療画像の信頼できない人間のアノテーションに基づくニューラルネットワークのトレーニング」 In Biomedical Imaging (ISBI 2018 , 2018 IEEE 15th International Symposium on .
0.79
IEEE, 2018, pp. 39–42.
IEEE, 2018, pp. 39-42。
0.43
[8] C. Xue, Q. Dou, X. Shi, H. Chen, and P. A. Heng, “Robust learning at noisy labeled medical images: Applied to skin lesion classification,” in ISBI, 2019.
C. Xue, Q. Dou, X. Shi, H. Chen, P. A. Heng, “Robust Learning at noisy labeled medical images: Applied to skin lesion classification” in ISBI, 2019。 訳抜け防止モード: [8]C.Xue,Q.Dou,X.Shi, H. Chen, P. A. Heng, “ノイズの多い医療画像のロバスト学習 : 皮膚病変分類への応用”。 ISBI、2019年。
0.83
[9] H. Zhu, J. Shi, and J. Wu, “Pick-and-learn: Automatic quality evaluation for noisy-labeled image segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Intervention.
H. Zhu, J. Shi, J. Wu, “Pick-and-learn: Automatic quality evaluation for noisy-labeled image segmentation” in International Conference on Medical Image Computing and Computer-Assisted Intervention。 訳抜け防止モード: [9]h.zhu、j.shi、j.wu 「ピック」と「学習」 : 雑音-ラベル付き画像セグメンテーションの自動品質評価」 international conference on medical image computing and computer - assisted intervention に参加して
0.70
Springer, 2019, pp. 576–584.
スプリンガー、2019年、p. 576-584。
0.54
[10] Y. Shu, X. Wu, and W. Li, “Lvc-net: Medical image segmentation with noisy label based on local visual cues,” in International Conference on Medical Image Computing and Computer-Assisted Intervention.
Y. Shu, X. Wu, W. Li, “Lvc-net: Medical image segmentation with noisey label based on local visual cues”. International Conference on Medical Image Computing and Computer-Assisted Intervention. (英語) 訳抜け防止モード: [10]y.shu、x. wu、w.li 「lvc-net : 局所視覚手がかりに基づくノイズラベルによる医用画像分割」 international conference on medical image computing and computer - assisted intervention に参加して
0.82
Springer, 2019, pp. 558–566.
春田、2019年、p.558-566。
0.56
[11] C. Xue, Q. Deng, X. Li, Q. Dou, and P.
[11]C.Xue,Q.Deng,X.Li,Q. Dou,P
0.34
-A. Heng, “Cascaded robust learning at imperfect labels for chest x-ray segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Intervention.
-A。 heng, “cascaded robust learning at imperfect labels for chest x-ray segmentation” in international conference on medical image computing and computer-assisted intervention” (英語) 訳抜け防止モード: -A。 heng氏: “胸部x線セグメンテーションの不完全ラベルにおけるロバスト学習のカスケード international conference on medical image computing and computer - assisted intervention に参加して
0.55
Springer, 2020, pp. 579–588.
スプリンガー、2020年、p. 579-588。
0.57
[12] J. Goldberger and E. Ben-Reuven, “Training deep neural-networks using
12] J. Goldberger と E. Ben-Reuven, “Deep Neural-networks のトレーニング
0.81
a noise adaptation layer,” in ICLR, 2017.
2017年、ICLRで「ノイズ適応層」を発表。
0.73
[13] G. Patrini, A. Rozza, A. K. Menon, R. Nock, and L. Qu, “Making deep neural networks robust to label noise: A loss correction approach,” in CVPR, 2017, pp. 2233–2241.
G. Patrini, A. Rozza, A. K. Menon, R. Nock, and L. Qu, “Making Deep Neural Network robust to label noise: A loss correct approach” in CVPR, 2017 pp. 2233–2241。 訳抜け防止モード: 13] g. patrini, a. rozza, a. k. menon, r. nock, そしてl. qu。 ラベルノイズにロバストなディープニューラルネットワーク cvpr, 2017 pp. 2233-2241における「損失補正アプローチ」。
0.70
[14] S. Zheng, P. Wu, A. Goswami, M. Goswami, D. Metaxas, and C. Chen, “Error-bounded correction of noisy labels,” in International Conference on Machine Learning.
14] s. zheng, p. wu, a. goswami, m. goswami, d. metaxas, c. chen, “error-bounded correction of noise labels”(ノイズラベルのエラー境界補正)は、機械学習に関する国際カンファレンスで発表された。 訳抜け防止モード: [14 ]S.Zheng,P.Wu,A.Goswa mi, M. Goswami, D. Metaxas, C. Chen, “Error - bounded correct of noisy labels”。 機械学習国際会議に参加して
0.85
PMLR, 2020, pp. 11 447–11 457.
pmlr, 2020, pp. 11 447-11 457。
0.72
[15] L. Jiang, Z. Zhou, T. Leung, L.
[15] L. Jiang, Z. Zhou, T. Leung, L.
0.49
-J. Li, and L. Fei-Fei, “MentorNet: Regularizing very deep neural networks on corrupted labels,” in ICML, 2018.
-j。 2018年、icmlで、liとl. fei-feiは“mentornet: regularizing very deep neural networks on brokened labels”を発表した。
0.67
[16] B. Han, Q. Yao, X. Yu, G. Niu, M. Xu, W. Hu, I. Tsang, and M. Sugiyama, “Co-teaching: Robust training of deep neural networks with extremely noisy labels,” in Advances in neural information processing systems, 2018, pp. 8527–8537.
[16] b. han, q. yao, x. yu, g. niu, m. xu, m. xu, w. hu, i. tsang, m. sugiyama, “co-teaching: robust training of deep neural networks with extreme noise labels” in advances in neural information processing systems, 2018, pp. 8527-8537. (英語) 訳抜け防止モード: [16]b.ハン,q.ヤオ,x.ユ, g. niu, m. xu, w. hu, i. tsang, and m. sugiyama, "co - teaching : robust training of deep neural network with ultra noise labels" (特集 深層ニューラルネットワーク) ニューラル・インフォメーション・プロセッシング・システムズ2018, pp. 8527-8537。
0.72
[17] D. Hendrycks, M. Mazeika, D. Wilson, and K. Gimpel, “Using trusted data to train deep networks on labels corrupted by severe noise,” in Advances in neural information processing systems, 2018, pp. 10 456– 10 465.
D. Hendrycks氏、M. Mazeika氏、D. Wilson氏、K. Gimpel氏は、ニューラル情報処理システムの進歩、2018年、pp. 10 456–10 465の中で、ラベルのディープネットワークをトレーニングするために信頼できるデータを使用します。
0.66
[18] J. Jiang, J. Ma, Z. Wang, C. Chen, and X. Liu, “Hyperspectral image classification in the presence of noisy labels,” IEEE Transactions on Geoscience and Remote Sensing, vol.
J. Jiang, J. Ma, Z. Wang, C. Chen, X. Liu, “Hyperspectral image classification in the presence of noisy labels”, IEEE Transactions on Geoscience and Remote Sensing, vol.
0.40
57, no. 2, pp. 851–865, 2018.
57, No. 2, pp. 851-865, 2018。
0.91
[19] X. Xia, T. Liu, N. Wang, B. Han, C. Gong, G. Niu, and M. Sugiyama, “Are anchor points really indispensable in label-noise learning?” in Advances in Neural Information Processing Systems, 2019, pp. 6838– 6849.
[19] x. xia, t. liu, n. wang, b. han, c. gong, g. niu, m. sugiyama, “アンカーポイントはラベルノイズ学習に本当に不可欠か? 訳抜け防止モード: [19 ]X. Xia, T. Liu, N. Wang, B. Han, C. Gong, G. Niu, M. Sugiyama ニューラル・インフォメーション・プロセッシング・システムにおける「アンカーポイントはラベルには本当に不可欠か?ノイズラーニング?」 2019 , pp . 6838 – 6849 .
0.84
英語(論文から抽出)
日本語訳
スコア
12 [42] E. Malach and S. Shalev-Shwartz, “Decoupling” when to update” from” how to update”,” in Advances in Neural Information Processing Systems, 2017, pp. 960–970.
12 [42] e. malach氏とs. shalev-shwartz氏は、"how to update"から"decoupling" when to update"を分離している。 訳抜け防止モード: 12 a b “42 ] malach and s. shalev - shwartz, “decoupling ” when to update ” from ” how to update ””[42 ] (英語) ニューラル・インフォメーション・プロセッシング・システムズ、2017年、p.960-970。
0.59
[43] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” arXiv preprint arXiv:2002.05709, 2020.
T. Chen, S. Kornblith, M. Norouzi, G. Hinton, “A simple framework for contrastive learning of visual representations” arXiv preprint arXiv:2002.05709, 2020。 訳抜け防止モード: [43 ]T. Chen, S. Kornblith, M. Norouzi, そしてG. Hintonは、“視覚表現の対照的な学習のためのシンプルなフレームワーク”だ。 arXiv preprint arXiv:2002.05709 , 2020。
0.75
[44] U. Avni, H. Greenspan, E. Konen, M. Sharon, and J. Goldberger, “Xray categorization and retrieval on the organ and pathology level, using patch-based visual words,” IEEE Transactions on Medical Imaging, vol.
U.Avni, H. Greenspan, E. Konen, M. Sharon, J. Goldberger, “X線分類と臓器と病理の検索は、パッチベースのビジュアルワードを使って行う”、IEEE Transactions on Medical Imaging, vol。
0.74
30, no. 3, pp. 733–746, 2010.
30 no. 3, pp. 733-746, 2010 頁。
0.82
[45] A. Tarvainen and H. Valpola, “Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results,” in Advances in neural information processing systems, 2017, pp. 1195–1204.
45] a. tarvainen, h. valpola, “mean teachers are better role models: weight-averaged consistency targets improve semi-supervised deep learning results” in advances in neural information processing systems, 2017 pp. 1195–1204。 訳抜け防止モード: [45 ]A. Tarvainen,H. Valpola, “平均教師はより良いロールモデルである。 In Advances in Neural Information Processing System, 2017, pp. 1195–1204。
0.55
[46] D. Berthelot, N. Carlini, I. Goodfellow, N. Papernot, A. Oliver, and C. A. Raffel, “Mixmatch: A holistic approach to semi-supervised learning,” in Advances in Neural Information Processing Systems, 2019, pp. 5049– 5059.
D. Berthelot, N. Carlini, I. Goodfellow, N. Papernot, A. Oliver, and C. A. Raffel, “Mixmatch: A holistic approach to semi-supervised learning” in Advances in Neural Information Processing Systems, 2019, pp. 5049–5059。 訳抜け防止モード: 46 ] d. berthelot, n. carlini, i. goodfellow, n. papernot, a. oliver, c. a. raffel, “mixmatch: a holistic approach to semi-supervised learning” ニューラル・インフォメーション・プロセッシング・システムズ, 2019, pp. 5049 – 5059。
0.77
[47] B. S. Veeling, J. Linmans, J. Winkens, T. Cohen, and M. Welling, “Rotation equivariant cnns for digital pathology,” in International Conference on Medical image computing and computer-assisted intervention.
[47]B.S. Veeling, J. Linmans, J. Winkens, T. Cohen, M. Welling, “Rotation equivariant cnns for digital pathology” in International Conference on Medical Image Computing and computer-assisted intervention。
0.46
Springer, 2018, pp. 210–218.
スプリンガー、2018年、p.210-218。
0.44
[48] X. Wang, Y. Peng, L. Lu, Z. Lu, M. Bagheri, and R. M. Summers, “Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2097–2106.
[48] x. wang, y. peng, l. lu, z. lu, m. bagheri, r. m. summers, “chestx-ray8: hospital-scale chest x-ray database and benchmarks on weakly supervised classification and localization of common thorax diseases” in the ieee conference on computer vision and pattern recognition, 2017 pp. 2097–2106 訳抜け防止モード: [48 ]X. Wang, Y. Peng, L. Lu, Z. Lu, M. Bagheri, R. M. Summers, “Chestx - ray8 : Hospital - scale chest x - ray database” 一般的な胸郭疾患の分類と局所化の弱さに関するベンチマークです。 In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2097–2106。
0.74
[49] A. Majkowska, S. Mittal, D. F. Steiner, J. J. Reicher, S. M. McKinney, G. E. Duggan, K. Eswaran, P.
A. Majkowska, S. Mittal, D. F. Steiner, J. J. Reicher, S. M. McKinney, G. E. Duggan, K. Eswaran, P.
0.45
-H. Cameron Chen, Y. Liu, S. R. Kalidindi et al , “Chest radiograph interpretation with deep learning models: assessment with radiologist-adjudica ted reference standards and population-adjusted evaluation,” Radiology, vol.
-h。 Cameron Chen, Y. Liu, S. R. Kalidindi et al , “Chhest radiograph interpretation with Deep Learning model: Assessment with Radioologist-adjudic ated Reference Standard and population-adjusted evaluation”. Radiology, vol. 訳抜け防止モード: -h。 cameron chen, y. liu, s. r. kalidindi et al. 深層学習モデルを用いた胸部x線画像の解釈 : 放射線科医による評価 - 基準基準と人口の調整-
0.72
294, no. 2, pp. 421–431, 2020.
294, no. 2, pp. 421-431, 2020。
0.89
[50] G. Nir, S. Hor, D. Karimi, L. Fazli, B. F. Skinnider, P. Tavassoli, D. Turbin, C. F. Villamil, G. Wang, R. S. Wilson et al , “Automatic grading of prostate cancer in digitized histopathology images: Learning from multiple experts,” Medical image analysis, vol.
50] g. nir, s. hor, d. karimi, l. fazli, b. f. skinnider, p. tavassoli, d. turbin, c. f. villamil, g. wang, r. s. wilson et al , "デジタル化病理組織画像における前立腺癌の自動段階的診断: 複数の専門家から学ぶこと", 医用画像解析, vol。 訳抜け防止モード: [50 ]G.Nir,S.Hor,D.Karimi , L. Fazli, B. F. Skinnider, P. Tavassoli, D. Turbin C. F. Villamil, G. Wang, R. S. Wilson et al デジタル化病理画像における前立腺癌の自動評価 複数の専門家から学ぶ”。 医用画像解析。
0.86
50, pp. 167–180, 2018.
50, pp. 167–180, 2018。
0.88
[51] D. Karimi, G. Nir, L. Fazli, P. C. Black, L. Goldenberg, and S. E. Salcudean, “Deep learning-based gleason grading of prostate cancer from histopathology images—role of multiscale decision aggregation and data augmentation,” IEEE journal of biomedical and health informatics, vol.
eee journal of biomedical and health informatics, vol. “51] d. karimi, g. nir, l. fazli, p. c. black, l. goldenberg, and s. e. salcudean, “深層学習に基づく前立腺がんの組織病理画像による評価 — マルチスケールの意思決定集約とデータ拡張の要領”、とieee journal of biomedical and health informatics, vol. は述べている。 訳抜け防止モード: [51 ]D.カリミ,G.ニル,L.ファズリ, P.C. Black、L. Goldenberg、S.E. Salcudean。 深層学習(Deep Learning) - 前立腺癌の病理像からのグリーソングレーディング(gleasongrading)に基づく。 IEEE Journal of Biomedical and Health Informationatics, vol。
0.69
24, no. 5, pp. 1413–1426, 2019.
24、no. 5, pp. 1413-1426、2019。
0.81
[52] D. Karimi, H. Dou, S. K. Warfield, and A. Gholipour, “Deep learning with noisy labels: Exploring techniques and remedies in medical image analysis,” Medical Image Analysis, vol.
52] d. karimi, h. dou, s. k. warfield, a. gholipour, “deep learning with noise labels: exploration techniques and remedies in medical image analysis”, medical image analysis, vol. 日本語版記事)。 訳抜け防止モード: [52 ]d. カリミ, h. dou, s. k. warfield, a. gholipour, “deep learning with noise labels” について 医療画像解析における技術と治療の探求」 医用画像解析
0.82
65, p. 101759, 2020.
65, p. 101759, 2020。
0.91
[53] S. K. Warfield, K. H. Zou, and W. M. Wells, “Simultaneous truth and performance level estimation (staple): an algorithm for the validation of image segmentation,” IEEE transactions on medical imaging, vol.
[53] s.k. warfield, k. h. zou, w. m. wells, “simultaneous truth and performance level estimation (staple: an algorithm for the validation of image segmentation”, ieee transactions on medical imaging, vol.)。 訳抜け防止モード: 53] s. k. warfield, k. h. zou, w. m. wells. 画像分割の検証のためのアルゴリズム「同時真理と性能レベル推定(staple)」 ieee transactions on medical imaging, vol. 参照。
0.77
23, no. 7, pp. 903–921, 2004.
23, No. 7, pp. 903–921, 2004。
0.47
[20] Y. Yao, T. Liu, B. Han, M. Gong, J. Deng, G. Niu, and M. Sugiyama, “Dual t: Reducing estimation error for transition matrix in label-noise learning,” Advances in Neural Information Processing Systems, vol.
20] y. yao, t. liu, b. han, m. gong, j. deng, g. niu, m. sugiyama, “dual t: reduce estimation error for transition matrix in label-noise learning” ニューラル情報処理システムにおける進歩。
0.82
33, 2020. [21] S. Sukhbaatar, J. Bruna, M. Paluri, L. Bourdev, and R. Fergus, “Training convolutional networks with noisy labels,” arXiv preprint arXiv:1406.2080, 2014.
33, 2020. S. Sukhbaatar, J. Bruna, M. Paluri, L. Bourdev, R. Fergus, “Training convolutional network with noisy labels, arXiv preprint arXiv:1406.2080, 2014”。 訳抜け防止モード: 33, 2020. [21]S. Sukhbaatar, J. Bruna, M. Paluri, L. BourdevとR. Fergus。 騒々しいラベルで畳み込みネットワークを訓練する。 arXiv preprint arXiv:1406.2080 , 2014
0.65
[22] D. Tanaka, D. Ikami, T. Yamasaki, and K. Aizawa, “Joint optimization
【22】田中d.、井上d.、山崎t.、愛沢k.『共同最適化』
0.56
framework for learning with noisy labels,” in CVPR, 2018.
とCVPR, 2018で述べている。
0.28
[23] M. Ren, W. Zeng, B. Yang, and R. Urtasun, “Learning to reweight
[23]M. Ren, W. Zeng, B. Yang, R. Urtasun, “Learning to reweight”
0.47
examples for robust deep learning,” in ICML, 2018.
と、2018年のicmlで述べている。
0.27
[24] D. T. Nguyen, C. K. Mummadi, T. P. N. Ngo, T. H. P. Nguyen, L. Beggel, and T. Brox, “Self: Learning to filter noisy labels with self-ensembling,” arXiv preprint arXiv:1910.01842, 2019.
D. T. Nguyen, C. K. Mummadi, T. P. Ngo, T. H. P. Nguyen, L. Beggel, T. Brox, “Self: Learning to filter noisey labels with self-ensembling” arXiv preprint arXiv:1910.01842, 2019”。 訳抜け防止モード: [24 ]D. T. Nguyen, C. K. Mummadi, T. P. N. Ngo, T. H. P. Nguyen, L. Beggel, T. Brox 「自己:自己でうるさいラベルをフィルターする学習」 arXiv preprint arXiv:1910.01842, 2019。
0.84
[25] J. Li, R. Socher, and S. C. Hoi, “Dividemix: Learning with noisy labels as semi-supervised learning,” arXiv preprint arXiv:2002.07394, 2020.
J. Li, R. Socher, S. C. Hoi, “Dividemix: Learning with noisy labels as semi-supervised learning” arXiv preprint arXiv:2002.07394, 2020。
0.46
[26] Y. Ding, L. Wang, D. Fan, and B. Gong, “A semi-supervised twostage approach to learning from noisy labels,” in 2018 IEEE Winter Conference on Applications of Computer Vision (WACV).
26] y. ding, l. wang, d. fan, b. gong, “a semi-supervised twostage approach to learning from noisy labels” 2018 ieee winter conference on applications of computer vision (wacv) で発表された。 訳抜け防止モード: [26 ]y ding, l. wang, d. fan, とb. gongは言う。“騒がしいラベルから学ぶための二段階的なアプローチ”だ。 2018年、ieee winter conference on applications of computer vision (wacv)。
0.73
IEEE, 2018, pp. 1215–1224.
IEEE, 2018, pp. 1215-1224。
0.85
[27] D. T. Nguyen, T.
[27] D. T. Nguyen, T。
0.46
-P. -N. Ngo, Z. Lou, M. Klar, L. Beggel, and T. Brox, “Robust learning under label noise with iterative noise-filtering,” arXiv preprint arXiv:1906.00216, 2019.
-P。 -N。 Ngo, Z. Lou, M. Klar, L. Beggel, T. Brox, “Robust learning under label noise with Iterative noise-filtering” arXiv preprint arXiv:1906.00216, 2019”。 訳抜け防止モード: -P。 -N。 Ngo, Z. Lou, M. Klar, L. Beggel そしてT. Brox氏は,“ラベルノイズ下でのロバストな学習と反復的なノイズフィルタリング”について語る。 arXiv preprint arXiv:1906.00216 , 2019
0.52
[28] S. Liu, J. Niles-Weed, N. Razavian, and C. Fernandez-Granda, “Earlylearning regularization prevents memorization of noisy labels,” Advances in Neural Information Processing Systems, vol.
28] s. liu, j. niles-weed, n. razavian, c. fernandez-granda, “早期学習の正規化はノイズラベルの記憶を妨げている”。 訳抜け防止モード: [28 ] S. Liu, J. Niles - Weed, N. Razavian, C. Fernandez - Granda, “Earlylearning regularization prevents memorization of noisy labels”。 ニューラル情報処理システムの進歩
0.83
33, 2020. [29] S. Min, X. Chen, Z.
33, 2020. [29]S. Min, X. Chen, Z.
0.45
-J. Zha, F. Wu, and Y. Zhang, “A two-stream mutual attention network for semi-supervised biomedical segmentation with noisy labels,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol.
-j。 zha, f. wu, y. zhang, “a aaai conference on artificial intelligence, vol. “a two-stream mutual attention network for semi-supervised biomedical segmentation with noise labels”(ノイズラベル付き半教師付きバイオメディカルセグメンテーションのための双方向相互注意ネットワーク)。 訳抜け防止モード: -j。 Zha, F. Wu, Y. Zhang両氏は,“ノイズラベル付きバイオメディカルセグメンテーションのための2つのストリーム相互注意ネットワーク”だ。 人工知能国際会議(AAAI)に参加して
0.74
33, no. 01, 2019, pp. 4578–4585.
33, no. 01, 2019, pp. 4578-4585。
0.90
[30] H. Le, D. Samaras, T. Kurc, R. Gupta, K. Shroyer, and J. Saltz, “Pancreatic cancer detection in whole slide images using noisy label annotations,” in International Conference on Medical Image Computing and Computer-Assisted Intervention.
30] h. le, d. samaras, t. kurc, r. gupta, k. shroyer, j. saltz, “pancreatic cancer detection in whole slide images using noise label annotations” は、医用画像コンピューティングとコンピュータ支援による介入に関する国際会議である。
0.83
Springer, 2019, pp. 541–549.
春田、2019年、p.541-549。
0.55
[31] Y. Wang, W. Liu, X. Ma, J. Bailey, H. Zha, L. Song, and S.
[31] Y. Wang, W. Liu, X. Ma, J. Bailey, H. Zha, L. Song, S。
0.47
-T. Xia, “Iterative learning with open-set noisy labels,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 8688–8696.
-T。 the ieee conference on computer vision and pattern recognition, 2018, pp. 8688-8696 で "オープンセットノイズラベルによる反復学習" が行われた。 訳抜け防止モード: -T。 Xia, “オープン-セットノイズラベルによる反復的な学習”。 IEEE Conference on Computer Vision and Pattern Recognition に参加して 2018 , pp . 8688–8696 .
0.64
[32] N. C. Codella et al , “Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic),” in Biomedical Imaging (ISBI 2018), 2018 IEEE 15th International Symposium on.
32] n. c. codella et al , “skin lesion analysis toward melanoma detection: a challenge at the 2017 international symposium on biomedical imaging (isbi) hosted by the international skin imaging collaboration (isic)” in biomedical imaging (isbi 2018), 2018 ieee 15th international symposium on。 訳抜け防止モード: 32 ] n. c. codella et al, “skin lesion analysis toward melanoma detection: a challenge at the 2017 international symposium on biomedical imaging (isbi)” 国際皮膚イメージングコラボレーション(isic)が主催する”in biomedical imaging(isbi 2018)”。 2018年ieee第15回国際シンポジウムに参加して
0.86
IEEE, 2018, pp. 168–172.
IEEE, 2018, pp. 168-172。
0.82
[33] S. Demyanov, R. Chakravorty, M. Abedini, A. Halpern, and R. Garnavi, “Classification of dermoscopy patterns using deep convolutional neural networks,” in 2016 IEEE 13th International Symposium on Biomedical Imaging (ISBI).
2016年 IEEE 13th International Symposium on Biomedical Imaging (ISBI) において、[33] S. Demyanov, R. Chakravorty, M. Abedini, A. Halpern, R. Garnavi, “Deep Convolutional Neural Network を用いた皮膚内視鏡パターンの分類” を報告した。
0.77
IEEE, 2016, pp. 364–368.
IEEE, 2016, pp. 364-368。
0.85
[34] J. Kawahara, A. BenTaieb, and G. Hamarneh, “Deep features to classify skin lesions,” in 2016 IEEE 13th International Symposium on Biomedical Imaging (ISBI).
[34] kawahara, a. bentaieb, g. hamarneh, “deep features to classification skin lesions” 2016 ieee 13th international symposium on biomedical imaging (isbi) で発表された。
0.76
IEEE, 2016, pp. 1397–1400.
IEEE, 2016, pp. 1397-1400。
0.85
[35] J. Premaladha and K. Ravichandran, “Novel approaches for diagnosing melanoma skin lesions through supervised and deep learning algorithms,” Journal of medical systems, vol.
J. Premaladha氏とK. Ravichandran氏は,“教師付き深層学習アルゴリズムによるメラノーマ皮膚病変の診断へのノーベルなアプローチ”を論じています。
0.69
40, no. 4, p. 96, 2016.
40,no.4,p.96,2016。
0.39
[36] L. Yu, H. Chen, Q. Dou, J. Qin, and P.
[36]L.Yu,H.Chen,Q.Dou,J. Qin,P。
0.73
-A. Heng, “Automated melanoma recognition in dermoscopy images via very deep residual networks,” IEEE transactions on medical imaging, vol.
Irshad, and A. H. Beck, “Deep learning for identifying metastatic breast cancer,” arXiv preprint arXiv:1606.05718, 2016.
Irshad, and A. H. Beck, “Deep Learning for identify metastatic breast cancer” arXiv preprint arXiv:1606.05718, 2016 訳抜け防止モード: Irshad, and A. H. Beck, “ディープラーニングによる転移性乳癌の同定” arXiv preprint arXiv:1606.05718 , 2016
0.85
[38] B. Kong, X. Wang, Z. Li, Q. Song, and S. Zhang, “Cancer metastasis detection via spatially structured deep network,” in International Conference on Information Processing in Medical Imaging.
[38] b. kong, x. wang, z. li, q. song, s. zhang, “cancer metastasis detection via spatially structured deep network” in international conference on information processing in medical imaging。 訳抜け防止モード: [38]b. kong, x. wang, z. li, q. song,s. zhang, “空間構造深層ネットワークによる癌転移の検出” international conference on information processing in medical imaging に参加して
0.83
Springer, 2017, pp. 236–248.
2017年、p.236-248。
0.55
[39] P. Courtiol, E. W. Tramel, M. Sanselme, and G. Wainrib, “Classification and disease localization in histopathology using only global labels: A weakly-supervised approach,” arXiv preprint arXiv:1802.02212, 2018.
P. Courtiol氏、E.W. Tramel氏、M. Sanselme氏、G. Wainrib氏は、「グローバルラベルのみを用いた病理組織における分類と疾病の局在化:弱い監督されたアプローチ」 arXiv preprint arXiv:1802.02212, 2018。
0.65
[40] Y. Liu, K. Gadepalli, M. Norouzi, G. E. Dahl, T. Kohlberger, A. Boyko, S. Venugopalan, A. Timofeev, P. Q. Nelson, G. S. Corrado et al , “Detecting cancer metastases on gigapixel pathology images,” arXiv preprint arXiv:1703.02442, 2017.
Y. Liu, K. Gadepalli, M. Norouzi, G. E. Dahl, T. Kohlberger, A. Boyko, S. Venugopalan, A. Timofeev, P. Q. Nelson, G. S. Corrado et al , “Detecting cancer metastass on gigapixel pathology images” arXiv preprint arXiv:1703.02442, 2017
0.47
[41] A. Cruz-Roa, H. Gilmore, A. Basavanhally, M. Feldman, S. Ganesan, N. N. Shih, J. Tomaszewski, F. A. Gonz´alez, and A. Madabhushi, “Accurate and reproducible invasive breast cancer detection in wholeslide images: A deep learning approach for quantifying tumor extent,” Scientific reports, vol.
41] a. cruz-roa, h. gilmore, a. basavanhally, m. feldman, s. ganesan, n. n. shih, j. tomaszewski, f. a. gonz ́alez, a. madabhushi, “wholeslide image: a deep learning approach for quantifying tumor extent: an quantifying tumor extent.”. 科学誌vol. 訳抜け防止モード: [41 ]A. Cruz-Roa, H. Gilmore, A. Basavanhally, M. Feldman, S. Ganesan, N. N. Shih, J. Tomaszewski F. A. Gonz ́alez と A. Madabhushi は、「クロスライド画像における正確な再現可能な乳がん検出 : 腫瘍範囲の定量化のための深層学習アプローチ」と述べている。 科学論文。