A Magnification-Aware Explainable EfficientNet-B3 Framework for Benign and Malignant Classification of Breast Histopathology Images

Abstract:

Breast cancer is diagnosed by examining stained tissue under a microscope, a task that depends on scarce specialists and on judgments that pathologists do not always share. We built a magnification-aware pipeline that first identifies the microscope magnification of an input image and then passes it to a dedicated EfficientNet-B3 classifier for benign versus malignant prediction. The models were fine-tuned from ImageNet weights on the BreakHis dataset (7,909 images from 82 patients) with class-weighted cross-entropy, and each prediction can be explained with Grad-CAM and LIME. On a held-out test set of 1,187 images, the four classifiers reached accuracies of 99.35% (40X), 97.57% (100X), 96.60% (200X) and 97.03% (400X), with ROC-AUC between 0.9928 and 0.9999. The magnification detector reached 98.06% accuracy. Pooled across magnifications, 15 of 815 malignant images were missed and 13 of 372 benign images were flagged as malignant. Published BreakHis studies report higher accuracy at several magnifications, so we make no claim of state-of-the-art accuracy. The contribution is a single deployable pipeline that combines routing, classification, two complementary explanations and a downloadable report in a Streamlit prototype. Three limits matter: the split was random at image level, so images from one patient may sit on both sides; no external dataset was used; and the explanations were inspected visually without pathologist annotation. The framework is a research prototype, not a diagnostic device.

Keywords: breast cancer, histopathology, BreakHis, EfficientNet-B3, Grad-CAM, LIME, explainable artificial intelligence,
magnification