BD-StableNet: a deep stable learning model with an automatic lesion area detection function for predicting malignancy...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep learning in medical imaging offers noninvasive breast cancer diagnosis.
• Expert annotation of large datasets is time-consuming and often infeasible.
• Lack of interpretability hinders deep learning adoption in medicine.

Purpose of the Study:

• To develop an interpretable deep stable learning model for detecting malignant tumors in BI-RADS category 3-4A breast lesions.
• To improve diagnostic accuracy and efficiency for early breast cancer detection.

Main Methods:

• Proposed BD-StableNet, a deep stable learning model for automatic lesion area detection.
• Utilized a retrospective dataset of 3103 breast ultrasound images from 493 patients.
• Compared BD-StableNet against mainstream deep learning models.

Main Results:

• BD-StableNet achieved high performance: accuracy=0.952, AUC=0.982, precision=0.970, recall=0.941, F1-score=0.955, specificity=0.965.
• Lesion area prediction and class activation maps confirmed model interpretability.
• Demonstrated superior prediction performance compared to other deep learning models.

Conclusions:

• BD-StableNet significantly enhances diagnostic accuracy and interpretability for BI-RADS 3-4A breast lesions.
• Offers a promising noninvasive approach for breast cancer diagnosis.
• Clinical use could reduce unnecessary biopsies and improve patient outcomes.