[Paper Review] Classification and Disease Localization in Histopathology Using Only Global Labels: A Weakly-Supervised Approach

Pierre Courtiol, Eric W. Tramel+2 more|arXiv (Cornell University)|Feb 1, 2018

AI in cancer detection32 references99 citations

TL;DR

This paper presents CHOWDER, a weakly supervised method for disease classification and localization in whole-slide histopathology images using only image-level labels, achieving competitive results without pixel-level annotations.

ABSTRACT

Analysis of histopathology slides is a critical step for many diagnoses, and in particular in oncology where it defines the gold standard. In the case of digital histopathological analysis, highly trained pathologists must review vast whole-slide-images of extreme digital resolution ($100,000^2$ pixels) across multiple zoom levels in order to locate abnormal regions of cells, or in some cases single cells, out of millions. The application of deep learning to this problem is hampered not only by small sample sizes, as typical datasets contain only a few hundred samples, but also by the generation of ground-truth localized annotations for training interpretable classification and segmentation models. We propose a method for disease localization in the context of weakly supervised learning, where only image-level labels are available during training. Even without pixel-level annotations, we are able to demonstrate performance comparable with models trained with strong annotations on the Camelyon-16 lymph node metastases detection challenge. We accomplish this through the use of pre-trained deep convolutional networks, feature embedding, as well as learning via top instances and negative evidence, a multiple instance learning technique from the field of semantic segmentation and object detection.

Motivation & Objective

• Address the high cost and scarcity of pixel-level annotations in histopathology image analysis.
• Develop a weakly supervised learning approach that localizes disease regions using only whole-slide labels.
• Leverage pre-trained CNN features with multiple instance learning to identify top indicative tiles and negative evidence.
• Produce interpretable localization maps that align with ground-truth regions without requiring segmentation masks.

Proposed method

• Use ResNet-50 features (2048-dim) extracted from 224x224 tiles as tile embeddings.
• Apply a one-dimensional embedding layer across tile indices to produce a consistent tile-level embedding.
• Retain top R and bottom R tile embeddings (top instances and negative evidence) for MIL-based classification.
• Classify with an MLP on the selected top/bottom embeddings to obtain slide-level predictions.
• Baseline aggregation methods (MaxPool, MeanPool) are compared to the proposed CHOWDER approach.
• Train via Adam with binary cross-entropy loss, employing ensembling (E networks) to improve generalization.

Experimental results

Research questions

• RQ1 Can disease localization in whole-slide histopathology images be achieved using only image-level labels (weak supervision) without pixel-level masks?
• RQ2 How does a MIL-based approach with top instances and negative evidence perform relative to standard pooling on heterogeneous WSI datasets?
• RQ3 To what extent can CHOWDER localize disease regions while matching or exceeding state-of-the-art weakly supervised methods on Camelyon-16 and TCGA datasets?

Key findings

• CHOWDER achieves competitive or superior AUC compared with strong-baseline pooling and WELDON across datasets.
• On Camelyon-16, CHOWDER with appropriate R produces higher AUC than baselines and performs comparably to human-level performance without using segmentation masks.
• On TCGA-Lung, CHOWDER shows robustness in a diffuse disease regime with AUC values around 0.90–0.92 for R in {1,10,100}.
• Localization maps produced by CHOWDER highlight tumor regions without pixel-level training data, closely aligning with ground-truth annotations in qualitative visualizations.
• Compared to the strongest published method on Camelyon-16, CHOWDER achieves substantial performance gains without ground-truth segmentation masks (AUC close to Bejnordi et al. 2017).
• Ensembling multiple CHOWDER models helps stabilize performance given small dataset sizes and limited hyperparameter tuning.