Thrilled to share my latest PhD work exploring fundamental questions about the transferability of multiple instance learning models in #computationalpathology, accepted as a spotlight paper to @icmlconf ! Many thanks to @richardjchen @GreatAndrew90 @AI4Pathology and all coauthors
📣 Excited to share our new ICML 2025 Spotlight article, “Do Multiple Instance Learning Models Transfer?” – addressing a foundational question for building robust and generalizable MIL models.
Read the article: arxiv.org/pdf/2506.09022
👉Enhanced Performance & Robustness: Pretrained MIL models consistently lead to improved performance even when the pre-training data comes from a different organ, site, disease model than the target task.
👉Aggregation Transfers: Transfer gains come from the MIL aggregation module, not just patch encoders. Resetting attention layers drops performance by 5–8%, showing they encode generalizable pooling logic.
👉Pancancer Generalization: Models pretrained on a more diverse and challenging data (e.g. 108-class pancancer classification task) achieve the stronger overall transfer performance.
👉Robust benefits across patch encoders: Benefits from MIL transfer are consistent across a wide range of patch encoders, from out-of-domain encoders such as ResNet50 pretrained on natural images, to in-domain encoders including Gigapath and UNIv2.
This research highlights supervised pretraining as a highly accessible path to generalizable MIL models, offering a data and compute-efficient route for developing slide level encoders with flexible combination of MIL method and patch encoder.
Congratulations to @Daniel__Shao @GreatAndrew90 @richardjchen and everyone else who contributed.
Stay tuned for an array of pre-trained MIL models ready to transfer to any task! Visit us at @icmlconf.View quoted post on X
