Deep Learning for Describing Breast Ultrasound Images with BI-RADS Terms.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer is the most common cancer in women, necessitating accurate diagnostic tools.
• Ultrasound interpretation requires expert knowledge, highlighting the need for computer-aided diagnosis (CAD) systems.
• The Breast Imaging-Reporting and Data System (BI-RADS) provides a standardized language for describing tumors and assessing malignancy.

Purpose of the Study:

• To develop and evaluate a deep neural network for automated tumor detection, description, and classification in breast ultrasound images.
• To enhance the explainability of AI-driven malignancy classification by integrating BI-RADS descriptors.
• To improve diagnostic accuracy and consistency in breast cancer assessment.

Main Methods:

• A deep neural network was trained on 749 expert-annotated nodules from public datasets.
• The YOLO detection algorithm was employed for Region of Interest (ROI) extraction.
• A multi-class classification model processed ROIs to output BI-RADS descriptors, BI-RADS classification, and malignancy prediction.

Main Results:

• The proposed model demonstrated superior agreement with expert radiologists (Cohen's kappa: 0.58 in cross-validation, 0.64 in testing) compared to other state-of-the-art CNNs.
• Integration of the YOLO algorithm significantly improved model performance.
• Training with BI-RADS descriptors enabled explainable malignancy classification without compromising accuracy.

Conclusions:

• The developed deep neural network effectively aids in breast cancer diagnosis by providing automated tumor analysis based on BI-RADS criteria.
• The system offers explainable AI, enhancing physician trust and understanding in the diagnostic process.
• This approach represents a significant advancement in computer-aided diagnosis for breast ultrasound, improving accuracy and interpretability.