ABMIL

ABMIL (Attention-Based Multiple Instance Learning) is a slide-level aggregation method that pools a whole-slide image's patch features into a single prediction using a learned attention mechanism, originally introduced by Ilse, Tomczak, and Welling at ICML 2018. In computational pathology it has become a standard downstream head: a patch encoder first turns a gigapixel slide into a bag of embeddings, and an ABMIL model weights and combines those embeddings for tasks such as cancer subtyping or biomarker prediction. This entry documents a specific set of pretrained ABMIL checkpoints released by the Mahmood Lab at Harvard Medical School and Brigham and Women's Hospital.

What distinguishes these checkpoints is not the architecture but the pretraining. Rather than training an ABMIL head from scratch on a small clinical cohort, the Mahmood Lab pretrained the aggregator on PC-108, a challenging 108-way pan-cancer morphological classification task spanning 24,000 whole-slide images from Mass General Brigham. The resulting weights are meant to be transferred and fine-tuned on downstream tasks, mirroring the transfer-learning practice long standard in NLP and computer vision but historically underused for MIL. The checkpoints accompany the study "Do Multiple Instance Learning Models Transfer?" (Shao et al., ICML 2025), which systematically evaluated MIL transferability.

Three variants are released, one per patch-feature backbone: abmil.base.uni_v2.pc108-24k (UNI2 features), abmil.base.conch_v15.pc108-24k (CONCH v1.5 features), and abmil.base.uni.pc108-24k (UNI features). Distributed through the Mahmood Lab MIL-Lab library under the FEATHER-24K label, they are downstream slide classifiers rather than self-supervised foundation models, and each is paired with the patch encoder whose embeddings it was trained to aggregate.

#Key Features

• Pan-cancer pretraining (PC-108): The aggregators are pretrained on a 108-way morphological classification task over 24,000 Mass General Brigham slides, yielding weights that transfer across organs and tasks instead of starting from random initialization.
• Three matched backbones: Separate checkpoints target UNI, UNI2, and CONCH v1.5 patch features, so users can pair the aggregator with whichever pathology patch encoder they already run.
• Attention-based pooling: A gated-attention mechanism assigns interpretable weights to individual patches before pooling, indicating which regions drive a slide-level prediction.
• Lightweight and fine-tunable: Each model has under one million parameters (roughly 0.79M for the CONCH v1.5 variant and 0.92M for the UNI variant), making fine-tuning feasible on a single consumer-grade GPU.
• Standardized loading: The MIL-Lab library exposes the checkpoints through a single create_model('abmil.base.<backbone>.pc108-24k', num_classes=...) call, with no attached classification head so users set the target class count.

#Technical Details

Each checkpoint is a base-size ABMIL network operating on pre-extracted patch embeddings; the CONCH v1.5 variant consumes 512x512 pixel patches at 20x magnification. The models were pretrained by supervised optimization on the PC-108 pan-cancer task and are released without a classification head, so downstream users attach and train a linear head for their own label set. The accompanying study assessed 11 MIL architectures across 21 pretraining tasks for morphological and molecular subtype prediction, and reported that pretrained MIL models consistently outperform scratch-trained ones even when pretrained on a different organ than the target. Pan-cancer PC-108 pretraining gave the largest transfer gains, improving average KNN performance by 9.8% and fine-tuning performance by 3.3% over random initialization, with ABMIL among the strongest aggregators.

#Applications

These checkpoints serve computational pathology researchers building slide-level classifiers on limited labeled data. Because clinical cohorts for a given biomarker or subtype are often small and weakly supervised, initializing an ABMIL head from PC-108 weights and fine-tuning offers a stronger starting point than training from scratch, particularly for cancer subtyping and molecular subtype prediction. The three backbone variants let teams reuse patch features they already extract with UNI, UNI2, or CONCH v1.5, and the models integrate with the Mahmood Lab Patho-Bench harness for standardized evaluation.

#Impact

These checkpoints operationalize a practical finding: MIL aggregators, like backbones in other domains, benefit from transfer learning, and pan-cancer pretraining can outperform much larger slide foundation models while using far less pretraining data. As sub-million-parameter downstream heads rather than self-supervised encoders, they sit at the boundary of the foundation-model landscape, but they lower the barrier to strong slide-level baselines. Access is restricted: the weights and MIL-Lab code carry a CC BY-NC-ND 4.0 research-only license and are gated behind registration, the PC-108 training corpus is private to a single US academic medical center, and no training code or intermediate checkpoints are released. The models have not been validated for clinical diagnostic use and require independent evaluation before any diagnostic application.