Multimodal Diagnosis of Breast Lesions Presenting as Architectural Distortion on DBT by Integrating with Ultrasound and MRI Features

Motivation: Diagnosis of architectural distortion (AD) on DBT is challenging, and supplementary features of breast Ultrasound (US) and MRI may help. Goal(s): To compare the diagnostic performance of models built using the radiological features of DBT, US, MRI, and the combined features. Approach: The BI-RADS categories of DBT, US and MRI were reported, and the machine learning (ML) algorithms were applied to build diagnostic models. Results: DBT, US, and MRI showed comparable diagnostic performance based on BI-RADS categories of 5, 4C, 4B, 4A, 3, 2. The accuracy and AUC were improved using the ML models developed with combined multimodal features. Impact: Diagnosis of lesions presenting as architectural distortion on Digital Breast Tomosynthesis (DBT) is difficult by all breast imaging modalities, and the performance can be improved with ML models developed using the combined multimodal features of DBT, US and MRI.