ODELIA ABMIL Model

Attention-Based Multiple Instance Learning (ABMIL) model for subject-level three-class breast MRI classification in the ODELIA Challenge 2025.

This model was contributed by the ABMIL team. The source implementation is available at adarshbhandaryp/ABMIL.

Method

The model represents each subject using 32 axial slices with pre-contrast, early post-contrast, and subtraction channels. A Swin Transformer Tiny backbone extracts slice features. Gated attention aggregates the slice features into a patient-level representation, followed by a three-class classifier. The source README reports focal loss with gamma 5.0.

Intended use and limitations

This release is for research and benchmarking only. It is not a medical device, is not clinically validated, and must not be used for diagnosis, prognosis, treatment, triage, or other clinical decisions. Performance may vary across institutions, scanners, acquisition protocols, and populations.

Setup

The source repository provides the complete training and evaluation code and requirements:

git clone https://github.com/adarshbhandaryp/ABMIL.git
cd ABMIL
pip install -r requirements.txt

Refer to the source repository for configuration-file usage and preprocessing requirements. The model expects ODELIA-style multi-sequence breast MRI input.

License and attribution

This repository is published under the ODELIA and ABMIL authors' permission.

The Swin Transformer implementation and pretrained ImageNet weights are third-party components. Their applicable licenses and terms must be preserved and reviewed before redistribution. See the dependency notices in the source repository and the relevant upstream package/model licenses.

Please cite or acknowledge:

• ODELIA Challenge 2025 and the ODELIA consortium.
• The ABMIL authors and source repository.
• The ODELIA paper: arXiv:2506.00474.