Development and evaluation of a deep incremental learning system for mammography image analysis

Breast cancer is a malignant tumor that mostly stems from ductal or lobular epithelial tissues. Each year, over two million people suffer from this disease and over half a million pass away globally, ranking breast cancer the most prevalent cancer type in women and the second cause of cancer-related deaths. Periodic screening programs such as mammography have significantly improved survival via early detection. Therefore, facilitating the interpretation for this modality through Computer-Aided Diagnosis (CAD) systems can accelerate specialists' decision-making and further raise the survival rates.

A CAD web-based system was designed through Unified Modeling Language (UML) according to the three-tier architecture. The design was implemented using C# programming language and the ASP.NET 8 framework along with the Microsoft Structured Query Language (MSSQL) database. The ViTGEMC deep ensemble model with Incremental Learning (IL) capabilities was incorporated as the prediction engine. Moreover, the system evaluation was conducted in two phases: a multi-reader and multi-case experiment to appraise the system's assistance effect, and the user experience assessment via the User Experience Questionnaire (UEQ).

The impact of software predictions was evaluated in an interventional test with the participation of four radiology specialists and 40 dual-view examinations, which yielded a 19.5% reduction in interpretation time (404.9 ± 104.2 vs. 325.8 ± 80.3 seconds, p = 0.0203) while maintaining accuracy (0.769 ± 0.024 vs. 0.800 ± 0.147, p = 0.6560). The usability assessment was further conducted by 10 experts in medicine, radiology, and health information technology fields, in which the system achieved scores of 1.4 to 2.3 across the six UEQ aspects of user experience.

The CAD software provides a convenient and enhanced environment which can aid radiologists in quicker diagnosis.