Development of an Explainable Deep Learning Model Based on Morphological Component Analysis (MCA) for Breast Cancer Detection in Ultrasound Images

Abstract

Purpose: This study aims to develop a deep learning model for breast ultrasound image analysis to improve lesion segmentation and support computer-aided breast cancer diagnosis. Design/Methodology/Approach: The study employs Morphological Component Analysis to decompose ultrasound images into structural, texture, and noise components during preprocessing. The resulting components are processed using an autoencoding convolutional network based on open optimisation. Grad-CAM is subsequently employed to provide visual explanations of the model’s decisions. The model was evaluated the model on the BUSI dataset, comprising 780 normal, benign, and malignant ultrasound images, using ten-fold cross-validation. Research Limitation: External validation on independent, larger clinical datasets is needed to assess generalizability and clinical applicability. Findings: The proposed model achieved a Dice Similarity Coefficient of 0.890, a Jaccard Index of 0.830, and a True Positive Rate of 0.906. The False Positive Rate and False Negative Rate were 0.096 and 0.094, respectively, while the Mean Absolute Error and Hausdorff Error were 4.189 and 21.071. Across five evaluated samples, the Dice score ranged from 87% to 96%, with an average of 92%. Combining structural and texture components produced more accurate segmentation than using individual components and baseline models. Grad-CAM maps also identified image regions influencing model decisions. Practical Implication: The model can support the development of accurate and transparent computer-aided breast cancer diagnosis systems. Social Implication: Improved and explainable ultrasound analysis may assist clinicians and contribute to more reliable breast cancer screening. Originality/Value: The study integrates morphological component analysis, convolutional learning, and explainable artificial intelligence into a unified model for breast ultrasound lesion segmentation.