Core Concepts
The BC-MRI-SEG benchmark provides a comprehensive evaluation of deep learning models for breast cancer tumor segmentation on publicly available MRI datasets, assessing their ability to generalize across diverse data sources.
Abstract
The BC-MRI-SEG benchmark consists of four public breast cancer MRI datasets, including ISPY1, BreastDM, RIDER, and DUKE, totaling 1,320 patients. The benchmark is designed to evaluate the performance and generalizability of deep learning models for binary breast cancer tumor segmentation.
The benchmark is divided into two stages:
Supervised training and evaluation using the ISPY1 and BreastDM datasets.
Zero-shot evaluation on the RIDER and DUKE datasets to assess the models' ability to generalize to unseen data.
Several state-of-the-art deep learning models are evaluated, including U-Net2D, U-Net3D, U-Net2.1D, SwinUNETR, Med-SA, and SegResNet. The results show that:
U-Net2.1D, a novel approach that leverages three neighboring MRI images, outperforms the traditional U-Net3D model in both supervised and zero-shot performance.
SegResNet, an asymmetrically large encoder architecture, achieves the best overall performance, particularly in the zero-shot evaluation.
Adapter-based fine-tuning, as used in Med-SA, fails to generalize well in the zero-shot setting.
The authors also provide an extensive list of publicly available breast cancer MRI datasets, which can be leveraged for further research and development of robust and generalizable models for clinical applications.
Stats
The ISPY1 dataset contains 67,080 images from 161 patients.
The BreastDM dataset contains 4,095 images from 232 patients.
The RIDER dataset contains 1,200 images from 5 patients.
The DUKE dataset contains 157,198 images from 922 patients.
Quotes
"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."