An interpretable CNN-based model for mass classification in mammography

Mammography is the primary screening method for lesion visualisation and detecting early potentially cancerous changes in breast tissue. The application of deep learning based computer-aided diagnosis (CADx) systems to mammography mass classification poses several challenges: confounding information is learned by a deep learning model, and it can be difficult for mammographic readers to understand how and why it makes a specific decision. In this work, we present a framework for interpretable convolutional neural network-based mammographic abnormality classification. In addition to predicting whether a mass lesion is benign or malignant, our work aims to follow the reasoning processes of mammographic readers in detecting clinically relevant semantic features, such as the shape characteristics of the mass. The framework includes model training that incorporates a combination of data with original images and data with pixel-wise annotations, leading to improved performance of the model. The proposed training method based on DenseNet121 achieved an improved accuracy of 86.1 ± 3.4% compared to 67.9 ± 3.8% for the original model on mass classification. The results show the proposed method highlighted the classification-relevant parts of the image, whereas the original method highlighted healthy tissue and confounding information. An interpretable algorithm is developed that explains the model using features representing specific clinical characteristics, thereby aiding in prediction. This allows mammographic readers to verify the model’s output for plausibility instead of relying on it blindly.