Abstract
In this study, a novel computer aided diagnosis (CADx) framework is devised to investigate interpretability for classifying breast masses. Recently, a deep learning technology has been successfully applied to medical image analysis including CADx. Existing deep learning based CADx approaches, however, have a limitation in explaining the diagnostic decision. In real clinical practice, clinical decisions could be made with reasonable explanation. So current deep learning approaches in CADx are limited in real world deployment. In this paper, we investigate interpretability in CADx with the proposed interpretable CADx (ICADx) framework. The proposed framework is devised with a generative adversarial network, which consists of interpretable diagnosis network and synthetic lesion generative network to learn the relationship between malignancy and a standardized description (BI-RADS). The lesion generative network and the interpretable diagnosis network compete in an adversarial learning so that the two networks are improved. The effectiveness of the proposed method was validated on public mammogram database. Experimental results showed that the proposed ICADx framework could provide the interpretability of mass as well as mass classification. It was mainly attributed to the fact that the proposed method was effectively trained to find the relationship between malignancy and interpretations via the adversarial learning. These results imply that the proposed ICADx framework could be a promising approach to develop the CADx system.
Introduction
While deep learning has achieved impressive accuracy in medical imaging, its clinical adoption is hindered by a lack of interpretability. In real-world practice, a radiologist provides a report based on the Breast Imaging Reporting and Data System (BI-RADS) standard, describing features like "spiculated margin" or "irregular shape." Most current Computer-Aided Diagnosis (CADx) systems provide a simple probability score without explaining the underlying logic. This "black box" nature makes it difficult for clinicians to trust AI in high-stakes decisions. The ICADx framework aims to bridge this gap by providing both a diagnosis and corresponding BI-RADS descriptions.
Methods
The core of the ICADx framework is a Generative Adversarial Network (GAN) consisting of two competing components: an Interpretable Diagnosis Network and a Synthetic Lesion Generative Network.
The diagnosis network is a multi-task CNN that simultaneously classifies malignancy, predicts mass margin and shape, and discriminates between real and synthetic images. The generative network, based on a U-Net architecture, re-synthesizes mass images based on specific malignancy conditions. These networks are trained in a minimax game: the generator creates realistic lesions to "fool" the diagnosis network, forcing the latter to become highly sensitive to subtle visual cues of margins and shapes.
In the test stage, a "Visual Interpreter" performs a sanity check. If the network predicts a mass is malignant due to a "spiculated" margin, the generator creates a synthetic version with those features. If the synthetic image closely matches the original, the decision is deemed interpretable and reliable.
Results and Discussion
The effectiveness of the proposed method was validated using 1,088 Regions of Interest (ROIs) from the Digital Database for Screening Mammography (DDSM).
The ICADx framework achieved an Area Under the Curve (AUC) of 0.908, significantly outperforming conventional CNNs (0.868) and standard multitask CNNs (0.893). Beyond classification accuracy, the model showed superior performance in identifying medical features, with shape prediction accuracy at 71.6% and margin prediction at 70.6%. These results suggest that forcing the model to learn medical descriptions through adversarial training actually enhances its diagnostic capabilities.
Conclusion
The ICADx framework successfully integrates interpretability into the CADx workflow by outputting BI-RADS terms and providing visual evidence through synthesized images. Future work will expand the feature scope to include other BI-RADS descriptors like density and explore 3D imaging modalities such as Digital Breast Tomosynthesis (DBT).
Summary updated