🤖 AI Summary
This study systematically investigates how pretrained feature extractors influence classification performance in whole-slide image (WSI) multi-instance learning (MIL). Addressing the interplay among backbone architecture, pretraining dataset, and pretraining paradigm, we conduct large-scale ablation studies across four state-of-the-art MIL models on TCGA-NSCLC and Camelyon16. Our results—quantitatively established for the first time—demonstrate that self-supervised pretraining (e.g., DINO) yields significantly greater performance gains than domain-specific supervised pretraining and is the dominant factor; vision transformers (ViTs) outperform CNNs, especially in deeper configurations; and pretraining on large-scale heterogeneous data substantially improves generalization. The work delivers a reproducible, evidence-based guideline for selecting WSI feature extractors, uncovers the root causes of current pathology foundation models’ superiority, and provides empirical foundations for designing efficient, scalable pathology large models.
📝 Abstract
Multiple instance learning (MIL) has become a preferred method for gigapixel whole slide image (WSI) classification without requiring patch-level annotations. Current MIL research primarily relies on embedding-based approaches, which extract patch features using a pre-trained feature extractor and aggregate them for slide-level prediction. Despite the critical role of feature extraction, there is limited guidance on selecting optimal feature extractors to maximize WSI performance. This study addresses this gap by systematically evaluating MIL feature extractors across three dimensions: pre-training dataset, backbone model, and pre-training method. Extensive experiments were conducted on two public WSI datasets (TCGA-NSCLC and Camelyon16) using four state-of-the-art (SOTA) MIL models. Our findings reveal that: 1) selecting a robust self-supervised learning (SSL) method has a greater impact on performance than relying solely on an in-domain pre-training dataset; 2) prioritizing Transformer-based backbones with deeper architectures over CNN-based models; and 3) using larger, more diverse pre-training datasets significantly enhances classification outcomes. We hope that these insights can provide practical guidance for optimizing WSI classification and explain the reasons behind the performance advantages of the current SOTA pathology foundation models. Furthermore, this work may inform the development of more effective pathology foundation models. Our code is publicly available at https://github.com/bryanwong17/MIL-Feature-Extractor-Selection