CMT-BUSNet: Adaptive Fusion-Based Triple-Branch Hybrid Architecture for Explainable Breast Ultrasound Tumor...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Accurate breast ultrasound tumor segmentation is crucial for diagnosis.
• Existing methods may lack boundary precision and interpretability.
• Hybrid architectures offer potential for improved performance.

Purpose of the Study:

• To introduce CMT-BUSNet, a novel hybrid architecture for breast ultrasound tumor segmentation.
• To evaluate CMT-BUSNet's performance against established benchmarks.
• To assess the model's explainability and cross-dataset transferability.

Main Methods:

• Developed CMT-BUSNet, integrating Convolutional Neural Networks (CNN), Mamba, and Transformer branches.
• Employed a CNN-anchored hierarchical parallel encoder with an Adaptive Feature Fusion Module (AFFM).
• Utilized Dense Nested Decoder and Boundary-Aware Composite Loss for segmentation and boundary delineation.

Main Results:

• Achieved a Dice Similarity Coefficient (DSC) of 0.9037 on the BUS-BRA dataset, outperforming nnU-Net v2 in boundary metrics (B-IoU: 0.611 vs. 0.557; HD95: 10.07 vs. 13.54 pixels).
• Demonstrated superior zero-shot transfer performance on the BUSI dataset (DSC = 0.6709 vs. 0.5579).
• Quantitative explainability (XAI) validation confirmed attribution faithfulness and uncertainty-error correlation.

Conclusions:

• CMT-BUSNet offers competitive segmentation accuracy with superior boundary delineation and built-in interpretability.
• The model shows promising preliminary cross-dataset transferability.
• Multicenter validation is recommended prior to clinical deployment.