Comparative Performance Analysis of Deep Learning Techniques for Automated Breast Cancer Classification from Medical Imaging

Abstract

Early and reliable breast cancer diagnosis remains a public-health priority, yet routine screening still suffers from inter-reader variability, heavy workload, and uneven resource availability. To move beyond generic claims of “deep learning works,” this paper reports a transparent, like-for-like comparison of three families of models under a single protocol and fixed pre-processing: a custom Convolutional Neural Network (CNN), MobileNetV3, and a Vision Transformer (ViT-L16). We evaluate on two complementary benchmarks—MIAS mammography (n=322 images) and BreakHis histopathology (n=7,909 tiles)—covering both binary (benign vs. malignant/abnormal) and multiclass settings. All models are trained on 224×224 inputs with harmonized augmentation; MIAS splits are image-level, whereas BreakHis splits are patient-level to avoid leakage. On MIAS, MobileNetV3 yields the strongest results with 93.25% accuracy for binary detection and 92.00% for multiclass classification, surpassing both the bespoke CNN and ViT-L16 under identical conditions. On BreakHis, ViT-L16 achieves the best validation performance (91.62% binary; 90.28% multiclass), which is consistent with the advantage of long-range self-attention for texture-rich histology. A side-by-side reading against recent literature shows our scores to be competitive and, in several cases, state-of-the-art under comparable validation protocols. Beyond headline numbers, we discuss deployment trade-offs: MobileNetV3 offers a favorable accuracy–efficiency balance for embedded or resource-constrained settings, while ViT-L16 is better suited to centralized, high-capacity pipelines. Taken together, the results argue for context-aware model selection and provide a reproducible baseline (unified splits, inputs, and training recipe) to facilitate independent verification and clinical translation.