[Development of an AI-based Positioning Technical Assistance System for Mammography]

We aimed to develop an AI-based system to score the positioning in mammography (MG), with the goal of establishing a foundation for future technical support.

Using 800 mediolateral oblique (MLO) images, we developed an AI model (Mask Generation Model) for automatic extraction of three regions: the pectoralis major muscle, the mammary gland region, and the nipple. Using this model, we extracted three regions from 1544 MLO images and generated mask images. The mask of the breast silhouette was automatically created using image processing techniques, resulting in a total of four mask images per image for 1544 MG. Afterward, we annotated those mask images and developed an AI model (Score Evaluation Model) for evaluating positioning with scores. The evaluation included six items: the insertion and shape of the pectoralis major muscle, the gap between the pectoralis major and the mammary gland, the extent of the mammary region, the breast silhouette, and the direction of stretching. We developed a system that quantifies MG positioning using AI-based scoring.

For the mask generation model, the DICE coefficient, which indicates the concordance of region extraction, was 0.968. For the score evaluation model, the accuracy of the six models was 0.825 or higher, and the AUC values from the ROC analysis were 0.873 or higher, indicating high performance.

A system that evaluates positioning with a score has been developed. The scores assigned to the specific items can be utilized as quantitative indicators for future technical support.