Binary breast cancer tumor segmentation with Magnetic Resonance Imaging (MRI) data is typically trained and evaluated on private medical data, which makes comparing deep learning approaches difficult. We propose a benchmark (BC-MRI-SEG) for binary breast cancer tumor segmentation based on publicly available MRI datasets. The benchmark consists of four datasets in total, where two datasets are used for supervised training and evaluation, and two are used for zero-shot evaluation. Additionally we compare state-of-the-art (SOTA) approaches on our benchmark and provide an exhaustive list of available public breast cancer MRI datasets.
To use our benchmark please refer to our GitHub. It will require downloading four datasets, using processing scripts to prepare the datasets, and then optionally downloading weights to replicate our results.
Additionally we form a list of publicly available breast cancer MRI data. Many contain multiple modalities which are listed in a table in the carousel below and available in the paper.
ISPY1, BreastDM, RIDER, DUKE, ISPY2, TCGA, DIAGNOSIS, QIN-Breast, NACT-Pilot,
Our experimental results suggest that adapter based tuning yields poor zero-shot performance, two neighboring images are sufficient for tumor segmentation, and an asymmetrically large encoder architecture outperforms a traditionally balanced encoder-decoder architecture.
The datasets, suitable for training larger medical models beyond our benchmark [46], offer opportunities such as utilizing pretext tasks [1]. Additionally, there is potential for finegrained labeling, including self-supervised learning for creating segmentation masks for DUKE [36] or enhancing masks in ISPY2 [27]. Further, extending our U-Net2.1D approach to more complex models and diverse research domains leveraging depth is an avenue for exploration.