Enhancing Accuracy in Breast Density Assessment Using Deep Learning: A Multicentric, Multi-Reader Study

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Oncology and Cancer Research

Background:

• Mammographic breast density is a key breast cancer risk factor.
• Current visual assessment by radiologists has significant interobserver variability.
• Inconsistent density assessment impacts breast cancer risk stratification.

Purpose of the Study:

• To develop and evaluate a deep learning-based automatic detection algorithm (DLAD) for automated breast density assessment.
• To compare the performance of DLAD against experienced radiologists.
• To improve the accuracy and consistency of breast density evaluation.

Main Methods:

• A multicentric dataset of 122 full-field digital mammography studies (488 images) was used.
• Ground truth was established by two experienced radiologists for 72 studies.
• DLAD performance was compared to five independent radiologists using accuracy, F1 score, precision, recall, and Cohen's Kappa.

Main Results:

• DLAD achieved an accuracy of 0.819 and a Cohen's Kappa of 0.708.
• The algorithm's performance matched or exceeded that of individual radiologists in several metrics.
• Statistical analysis showed no significant difference in accuracy between DLAD and radiologists.

Conclusions:

• The DLAD demonstrates robust and competitive performance in breast density assessment.
• Automated evaluation using DLAD can enhance accuracy and consistency compared to manual methods.
• This algorithm offers a reliable tool to improve breast cancer screening outcomes.