Lesion-Context Spectral Prototype Learning for Cross-Dataset Benign-Malignant Classification of MIAS and DDSM Mammograms

Abstract

Mammogram classification often obtains high within-dataset accuracy while remaining sensitive to acquisition style, background artefacts, and the size of the cropped lesion. This study proposes a lesion-context spectral prototype network (LCSP-Net) for benign-malignant classification using MIAS and DDSM mammograms. The method first removes non-breast background, performs robust percentile normalization and local contrast enhancement, and constructs paired lesion and surrounding-context crops from the available abnormality annotations. Three complementary representations are then learned: a lesion representation, a wider contextual representation, and a high-frequency wavelet representation. A data-dependent gate fuses these features, while a cross-dataset prototype term encourages benign and malignant embeddings from MIAS and DDSM to remain close within the same class and separated across classes. This design makes the contribution reside in the learning objective and evidence fusion rather than in a post-hoc attention map. Grad-CAM is therefore used only for interpretation after prediction. A patient-wise evaluation protocol is specified to prevent leakage between training and test images. Preliminary aggregate values supplied for this manuscript show accuracy increasing from 96.15% for the lesion CNN to 99.14% for the full ablation sequence; however, these values must be verified from locked raw predictions before submission. The reconstructed confusion matrix and illustrative Grad-CAM panel are explicitly marked as provisional and are not treated as independent experimental evidence. The resulting framework provides a reproducible route for testing whether lesion evidence, local tissue context, spectral detail, and cross-dataset representation consistency improve mammogram classification beyond single-stream CNN or transformer baselines.