An explainable and comprehensive BI-RADS assisted diagnosis pipeline for mammograms

Abstract

Background and Objective:

The growing adoption of computer-aided diagnosis systems is transforming cancer diagnostics in mammography, but it still needs improvement in comprehensive diagnostics, early screening accuracy, and system explainability. To achieve these, we introduce an explainable pipeline designed to provide comprehensive diagnostic suggestions according to the Breast Imaging Reporting and Data System (BI-RADS) lexicon.

Methods:

The proposed pipeline employs a multi-task framework, with BI-RADS assessments as the main task, complemented by the classification of masses, calcifications, architectural distortions, and breast density. We develop the Multi-scale Feature Fusion with Spatial Attention (MFFSA) module and the Low-level Feature Spatial Attention (LLSA) block. The MFFSA module extracts features from small lesions by fusing multi-scale features. The LLSA block makes the fusion process focus on the details of lesions. The system’s explainability is enhanced by two explainable methods, which clarify the effectiveness of the LLSA block in detecting small lesions and reveal how multi-scale features impact predictions in the MFFSA module. It also connects the prediction results of BI-RADS assessment with medical indicators.

Results:

Experiments on open-source mammography datasets of CBIS-DDSM, CDD-CESM and InBreast show the proposed method achieves an AUC of 91.9% (95% CI: 90.5%–93.2%), 91.8% (95% CI: 90.8%–92.8%) and 97.0% (95% CI: 94.7%–98.9%) respectively for the BI-RADS assessment task, and over 87% across all tasks. These results highlight the state-of-the-art performance and the pipeline’s ability to provide precise and comprehensive diagnostic suggestions.

Conclusions:

The proposed pipeline represents an advancement in applying artificial intelligence to mammography. Code available at https://github.com/jjjjjjjjj58/BI-RADS-Diagnosis-for-Mammograms.