Multi-site validation of an interpretable model to analyze breast masses

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep learning models show promise in breast lesion malignancy prediction.
• Interpretability in AI models is crucial for clinical trust and adoption.
• External validation is essential to assess model generalizability.

Purpose of the Study:

• To perform the first external validation of IAIA-BL, a deep learning-based, inherently interpretable breast lesion malignancy prediction model.
• To assess the generalizability of both performance and interpretability of IAIA-BL on external patient cohorts.
• To compare IAIA-BL against black-box baseline models in external validation settings.

Main Methods:

• External validation of the IAIA-BL model on two independent patient datasets (iCAD and Emory University).
• Evaluation of mass margin and malignancy classification performance using Area Under the Curve (AUC).
• Assessment of model interpretability through analysis of model activation on relevant lesion areas.

Main Results:

• IAIA-BL exhibited decreased mass margin classification performance on external datasets compared to the internal dataset, as measured by AUC.
• Malignancy classification performance also showed a reduction, but AUC 95% confidence intervals overlapped across all sites.
• Model interpretability, indicated by activation on relevant lesion portions, was maintained across all tested patient populations.

Conclusions:

• Model interpretability demonstrated generalizability across external datasets, even when classification performance experienced a slight decline.
• This study highlights the potential for interpretable AI models in breast cancer diagnostics to maintain transparency in real-world clinical settings.
• The findings support the development of inherently interpretable AI tools for reliable breast lesion assessment.