Retrospective multi-reader evaluation of a Deep-Learning-based cancer detection model in Contrast-Enhanced Mammography.

OBJECTIVES: To evaluate the clinical impact of a deep-learning-based computer-aided diagnostic tool applied to contrast-enhanced mammography on radiologists' diagnostic accuracy and reading efficiency. METHODS: In this retrospective single-centre study, 150 contrast-enhanced mammography examinations were interpreted by five radiologists. Each case was read and annotated twice in two sequential sessions: first without assistance and then with the computer-aided diagnostic tool, with a minimum 4-week washout period. Diagnostic performance was assessed at lesion and breast levels using free-response receiver operating characteristic analysis, area under the curve, and false-positive rate. Reading time for one radiologist was recorded as an exploratory measure. RESULTS: The computer-aided diagnostic tool was associated with improved lesion-level performance for most radiologists, with a mean increase in the area under the free-response curve of 0.030 ± 0.018. At 89% sensitivity, false-positive findings decreased by an average of 11.4% ± 24% across all radiologists. No improvement was observed at the breast level, possibly related to the tool's single-view analysis and dataset characteristics. An exploratory analysis showed a 31% reduction in reading time for one radiologist. CONCLUSION: Integration of a deep learning-based computer-aided diagnostic tool into contrast-enhanced mammography was associated with improved lesion-level diagnostic performance and reduced false-positive findings, without measurable improvement at the breast level. ADVANCES IN KNOWLEDGE: To our knowledge, this study provides one of the early clinical evaluations of a deep learning-based computer-aided diagnostic tool for contrast-enhanced mammography, suggesting potential improvements in lesion-level diagnostic performance and workflow efficiency while highlighting current limitations at the breast level.