Rahul Sawhney

Master's Student

Project Topic: Weak-Supervision for Breast DCE-MRI Segmentation via CAM/MIL Pseudo-Masks
Supervisors: Erik Gösche

Description

I am benchmarking weak-supervision strategies for tumour segmentation on the MAMA-MIA DCE-MRI cohort to quantify the Dice, HD95, and MSD trade-offs against full voxel supervision. I first reproduce the official nnU-Net baseline so the comparison uses identical preprocessing and splits. I then train a CAM/MIL classifier on breast-level response labels, convert its attention maps into pseudo-masks, and validate them against expert annotations. Those pseudo-masks feed a weakly supervised nnU-Net variant so I can measure the accuracy gap versus the baseline. I analyse error distribution, sample efficiency, and computational cost to understand when weak labels are clinically viable. If time permits, I will test the pipeline’s transferability on the multi-centre ODELIA dataset.

References

[1] Lidia Garrucho et al., “A large-scale multicenter breast cancer DCE-MRI benchmark dataset with expert segmentations,” *Scientific Data*,12:453, 2025. DOI: 10.1038/s41597-025-04707-4.

[2] Maximilian Ilse, Jakub Tomczak, and Max Welling, “Attention-based Deep Multiple Instance Learning,” *International Conference on Machine Learning*, 2018.

[3] Fabian Isensee et al., “nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation,” *Nature Methods*, 18, 2036–2047, 2021. DOI: 10.1038/s41592-020-01008-z.