[논문 리뷰] Semi-Supervised Histology Classification using Deep Multiple Instance Learning and Contrastive Predictive Coding

Ming Y. Lu, Richard J. Chen외 3명|arXiv (Cornell University)|2019. 10. 23.

AI in cancer detection참고 문헌 21인용 수 33

한 줄 요약

두 단계 반지도 학습 파이프라인 (CPC 사전학습 + attention-based MIL) 은 한정된 라벨과 선택적 인코더 고정으로 메모리 절약을 달성하면서 BACH의 이진 유방암 조직학 분류에서 최첨단 성능을 달성합니다.

ABSTRACT

Convolutional neural networks can be trained to perform histology slide classification using weak annotations with multiple instance learning (MIL). However, given the paucity of labeled histology data, direct application of MIL can easily suffer from overfitting and the network is unable to learn rich feature representations due to the weak supervisory signal. We propose to overcome such limitations with a two-stage semi-supervised approach that combines the power of data-efficient self-supervised feature learning via contrastive predictive coding (CPC) and the interpretability and flexibility of regularized attention-based MIL. We apply our two-stage CPC + MIL semi-supervised pipeline to the binary classification of breast cancer histology images. Across five random splits, we report state-of-the-art performance with a mean validation accuracy of 95% and an area under the ROC curve of 0.968. We further evaluate the quality of features learned via CPC relative to simple transfer learning and show that strong classification performance using CPC features can be efficiently leveraged under the MIL framework even with the feature encoder frozen.

연구 동기 및 목표

• Address overfitting and limited labeled data in deep MIL for histology classification.
• Leverage self-supervised feature learning to learn rich representations from unlabeled patches.
• Integrate CPC with MIL to improve performance while enabling memory-efficient training.
• Demonstrate performance gains and analyze feature utility under frozen vs. fine-tuned encoders.

제안 방법

• Use attention-based MIL to aggregate patch embeddings into bag representations for image-level classification.
• Pretrain the feature encoder on unlabeled patches via Contrastive Predictive Coding (CPC) to learn histology-specific features.
• Apply a smooth margin-based loss with KL-divergence regularization on negative bags to prevent overfitting to few informative instances.
• Experiment with ImageNet transfer learning vs CPC pretraining, with encoder frozen or finetuned, under MIL.
• Use a modified ResNet50 encoder and a compact gated attention-MIL network for bag prediction.
• Evaluate on five random splits of the BACH dataset (breast cancer histology) with 25% validation.

실험 결과

연구 질문

• RQ1 Can CPC-based self-supervised pretraining improve MIL-based histology classification when labeled data are scarce?
• RQ2 Does freezing the encoder after CPC pretraining affect MIL performance compared to finetuning?
• RQ3 How does the proposed smooth SVM loss with KL-divergence regularization influence negative bag handling in MIL?
• RQ4 What are the comparative gains of CPC pretraining versus ImageNet transfer learning within the MIL framework for histology slides?

주요 결과

Method	Accuracy (%)	AUC ROC
MIL + ImageNet (CE)	84.4 ± 9.40	0.933 ± 0.514
MIL + ImageNet (R)	86.0 ± 4.64	0.939 ± 0.240
MIL + CPC (CE)	91.8 ± 7.53	0.959 ± 0.052
MIL + CPC (R)	95.0 ± 2.65	0.968 ± 0.022
MIL	62.6 ± 11.6	0.611 ± 0.186
MIL + ImageNet	86.0 ± 4.64	0.939 ± 0.024
MIL + CPC	95.0 ± 2.65	0.968 ± 0.022
MIL + ImageNet, Frozen	82.8 ± 2.95	0.891 ± 0.026
MIL + CPC, Frozen	90.6 ± 2.88	0.939 ± 0.024

• CPC pretraining + MIL with smooth SVM loss + KL-div regularization achieves the highest mean accuracy and AUC (95.0% ±2.65, 0.968 ±0.022) among compared methods.
• CPC pretraining outperforms ImageNet transfer learning across evaluated setups, both when encoder is frozen and when finetuned.
• MIL with CPC (frozen encoder) still yields strong performance (90.6% ±2.88 accuracy, 0.939 ±0.024 AUC).
• Using a frozen encoder reduces trainable parameters to under 800k, enabling memory-efficient training on large bags.
• MIL alone performs poorly on this dataset, highlighting the benefit of CPC pretraining for feature learning in weakly supervised histology.
• Training with the smooth SVM loss + KL-div regularization consistently improves results over cross-entropy across splits.