Abstract
<title>Abstract</title> <p>Breast ultrasound imaging is widely used to support breast lesion assessment, but automated classification remains challenging because ultrasound images can be noisy, operator-dependent, and visually variable. This paper presents a preliminary convolutional neural network (CNN) baseline for breast ultrasound classification using MATLAB. The study uses the Breast-Lesions-USG (BrEaST) dataset, a public breast ultrasound dataset containing 256 scans with clinical descriptors, image labels, masks, BI-RADS-related fields, and histopathological or follow-up verification. The implemented workflow reads ultrasound images and clinical metadata, creates a binary malignant versus non-malignant label from the dataset classification field, resizes images to 224 × 224 pixels, applies normalization and augmentation, and trains a CNN using the Adam optimizer. The final comparative experiment reported an improvement in test accuracy from 70.88% to 78.43%, with precision increasing from 0.600 to 0.625. The exploratory IDC/ILC component was derived from text matching within the diagnosis field; therefore, it is reported as label exploration rather than as a validated subtype classifier. No explainability method was applied to the reported CNN model. A separate exploratory LIME-style decision-tree analysis is retained in the supplementary code for transparency, but it is not reported as a validated CNN explainability result. Instead, explainability is presented as the next research direction, motivated by the need for clinically interpretable breast imaging AI. The study is intended as a first documented historical baseline that can be strengthened through fixed-seed reruns, external validation, Grad-CAM-style visual explanations, uncertainty estimation, and future multimodal image-text modelling.</p>