Evidential learning driven Breast Tumor Segmentation with Stage-divided Vision-Language Interaction
New AI uses text prompts and uncertainty quantification to tackle low-contrast MRI scans where tumors blend with healthy tissue.
A research team led by Jingxing Zhong and Qingtao Pan has introduced TextBCS, a new AI model designed to solve a critical problem in medical imaging: accurately outlining breast tumors in MRI scans. Traditional deep learning methods often struggle because tumors can have low visual contrast with surrounding healthy tissue and blurry boundaries. TextBCS innovates by integrating a 'stage-divided vision-language interaction,' where descriptive text prompts guide the visual analysis at every step of the image processing pipeline. This allows the model to leverage contextual information to pinpoint lesion areas that are visually ambiguous.
Beyond just finding the tumor, TextBCS incorporates 'evidential learning' to tackle uncertainty. Instead of just outputting a segmentation mask, it quantifies how confident it is in each pixel's classification, especially at fuzzy tumor boundaries. It does this by modeling the probability distribution using a variational Dirichlet method. This gives radiologists a confidence map alongside the segmentation, highlighting areas where the AI is less certain and human review is most needed. Extensive testing shows TextBCS achieves state-of-the-art segmentation performance on public breast MRI datasets, demonstrating its potential as a powerful assistive tool for improving diagnostic accuracy and consistency in oncology.
- Integrates text prompts with MRI visuals at every processing stage to locate low-contrast tumors.
- Uses evidential learning to output confidence scores, quantifying uncertainty at blurry tumor boundaries.
- Outperforms existing segmentation networks on public datasets, showing superior accuracy.
Why It Matters
This could lead to earlier, more precise breast cancer detection and better-targeted treatment plans, directly impacting patient outcomes.