Self-Supervision Enhances Instance-based Multiple Instance Learning Methods in Digital Pathology: A Benchmark Study

TL;DR

The paper addresses WSI classification via MIL, showing that, with strong SSL-derived features, simple instance-based MIL can rival or surpass complex embedding-based MIL methods. It conducts a comprehensive benchmark across four histopathology datasets, evaluating 10 MIL strategies, 6 SSL methods, 4 backbones, and 4 foundation models, totaling 710 configurations, while introducing four new instance-based pooling operators. Key findings demonstrate that instance-based MIL paired with robust SSL not only achieves competitive performance but also attains state-of-the-art results on BRACS and Camelyon16, with superior interpretability for clinicians. The work highlights the practical impact of prioritizing pathology-aware SSL development and simpler, more interpretable MIL designs over increasingly complex embedding-based approaches, and provides public code and models to accelerate future research.

Abstract

Multiple Instance Learning (MIL) has emerged as the best solution for Whole Slide Image (WSI) classification. It consists of dividing each slide into patches, which are treated as a bag of instances labeled with a global label. MIL includes two main approaches: instance-based and embedding-based. In the former, each patch is classified independently, and then the patch scores are aggregated to predict the bag label. In the latter, bag classification is performed after aggregating patch embeddings. Even if instance-based methods are naturally more interpretable, embedding-based MILs have usually been preferred in the past due to their robustness to poor feature extractors. However, recently, the quality of feature embeddings has drastically increased using self-supervised learning (SSL). Nevertheless, many authors continue to endorse the superiority of embedding-based MIL. To investigate this further, we conduct 710 experiments across 4 datasets, comparing 10 MIL strategies, 6 self-supervised methods with 4 backbones, 4 foundation models, and various pathology-adapted techniques. Furthermore, we introduce 4 instance-based MIL methods never used before in the pathology domain. Through these extensive experiments, we show that with a good SSL feature extractor, simple instance-based MILs, with very few parameters, obtain similar or better performance than complex, state-of-the-art (SOTA) embedding-based MIL methods, setting new SOTA results on the BRACS and Camelyon16 datasets. Since simple instance-based MIL methods are naturally more interpretable and explainable to clinicians, our results suggest that more effort should be put into well-adapted SSL methods for WSI rather than into complex embedding-based MIL methods.

Figures (12)

• Figure 1: Self-supervision (SSL) closes the gap between instance- and embedding-based multiple-instance learning (MIL) methods in whole slide image (WSI) classification. Instance-based MIL methods, when combined with robust SSL feature extractors, are on par or even outperform complex, state-of-the-art embedding-based methods in WSI classification. The Y-axis represents the AUC scores of different MIL methods for four backbones (Vit_tiny, ViT_Small, ResNet18 and ResNet50) on the Camelyon16 dataset (at a resolution 20x for ResNet50 and 10x for the other backbones). For each backbone and MIL method, we show the box-plot of 3 self-supervised pre-trainings: DINO caron_emerging_2021, MOCO-V3 he_momentum_2020 and Barlow Twins zbontar_barlow_2021.
• Figure 2: Pipeline of the WSI classification. At the top, a) represents a self-supervised pre-training of an encoder f on patches from all slides. Once pre-trained, the encoder f extracts a vector of features (i.e., representation) per patch. Representations of patches of the same slide are then stacked together. At the bottom, b.1) Embedding-based MIL: features from the same slide are aggregated together using a feature aggregator $\sigma$ into a bag-level representation. A bag/slide classifier g is then trained on bag/slide representations to predict the slide label. b.2) Instance-based MIL: an instance classifier h assigns scores (i.e., class probability) to each patch, which are then pooled using a pooling mechanisms $\sigma$ to predict the slide label.
• Figure 3: Visual explanation of the proposed validation pipeline.
• Figure 4: ROC curves for 12 MIL methods, features extracted by ResNet18 backbone pretrained with Barlow Twins. The curves are plotted across three datasets—BRACS, TCGA-NSCLC, and Camelyon16. The best-performing method on each dataset is highlighted in a black curve with a larger line width.
• Figure 5: Grid Search of Learning Rate on Validation set of Camelyon16 dataset with 6 pre-training methods and 12 MILs with ResNet50 backbone