The Art of Perception: Cross-Modality Learning for Accurate Tumour Segmentation in Breast Imaging

Unlocking the potential of advanced imaging techniques, this study proposes an improved version of TransUNet, which includes a Multi-Scale Attention Module and Cross-Modality Adaptation, enabling the unleashing of advanced imaging for the segmentation of tumors in the breast. Our model is designed to cope with the challenges related to DBT, MRI, and ultrasound, capturing both fine and large-scale tumor features while generalizing seamlessly to multiple imaging modalities. The experimental results have shown that the performance of the proposed model outperforms all state-of-the-art methods with a Dice score of 65.50%, mIoU of 60.30%, and HD95 of 6.00 mm, while it still keeps computationally efficient: 25.40G FLOPs and 43.50M parameters, which makes it very promising for large-scale clinical deployment.