An Explainable Clinical Decision Support System for Breast Cancer Diagnosis Using Ultrasound Imaging and Clinical Feature Analysis

Abstract

Breast cancer is one of the leading causes of cancer-related mortality among women worldwide, highlighting the importance of accurate and early diagnosis. Among the available imaging techniques, ultrasound is widely used because it is cost-effective, non-invasive, and does not expose patients to ionizing radiation. Although many computer-aided diagnosis (CAD) systems have been developed to assist clinicians, most existing approaches are designed to perform only a single task, such as lesion segmentation or tumor classification. In addition, these systems often provide limited explanations for their predictions, reducing their usefulness in supporting clinical decision-making. This paper presents an Explainable Clinical Decision Support System (CDSS) that integrates ultrasound image analysis with structured clinical feature assessment within a unified software platform. The proposed framework combines deep learning, radiomics, and machine learning techniques to support breast cancer diagnosis using ultrasound images and structured clinical information. It employs a U-Net model to accurately segment breast lesions, followed by a ResNet18-based classifier to distinguish between benign and malignant cases. Radiomics features extracted from the segmented lesions, together with structured clinical attributes, are used by Random Forest models to predict the status of estrogen receptor (ER), progesterone receptor (PR), and HER2 biomarkers. Rather than predicting hormone receptor (HR) status directly, the system determines it from the predicted ER and PR results in accordance with standard clinical practice. To improve transparency and increase clinicians' confidence in the model's decisions, Grad-CAM is incorporated to highlight the image regions that contribute most to each classification. The complete framework is implemented as an interactive Streamlit application that supports patient registration, clinical data entry, ultrasound image upload, automated diagnosis, Grad-CAM visualization, patient record management, analytical dashboards, and PDF report generation. In addition, the application records several clinically relevant parameters, including symptom duration, breast quadrant involvement, hormonal therapy history, and histopathology findings, which are often unavailable in existing diagnostic systems. The proposed approach was evaluated using the BUSI breast ultrasound dataset, where the ResNet18 classifier achieved an overall accuracy of 96.80%. It also obtained a macro precision of 96.52%, macro recall of 96.74%, and a macro F1-score of 96.60%, demonstrating reliable classification performance across different categories. For lesion segmentation, the U-Net model achieved a Dice Similarity Coefficient (DSC) of 0.9124 and an Intersection over Union (IoU) of 0.8468, indicating accurate lesion localization suitable for both visual interpretation and radiomics feature extraction. To assess the generalizability of the framework, an independent external dataset containing 20 breast ultrasound cases with corresponding clinical information from Siddaganga Medical College & Research Institute was used exclusively for validation. Since the imaging data and clinical records originated from different sources and were not matched at the individual patient level, the imaging and clinical prediction models were trained independently and integrated into a unified clinical decision support system rather than performing true patient-level multimodal fusion. Overall, the proposed system provides an interpretable and comprehensive decision support framework for breast cancer diagnosis by combining ultrasound image analysis with structured clinical information, enabling accurate prediction, enhanced explainability, and practical support for clinical decision-making.