MORANet: A Modality-Oriented Representation and Attention Network for Breast Cancer Diagnosis
Contributors
Himanish Shekhar Das
Dr. Subrata Chowdhury
Keywords
Proceeding
Track
General Track
License
Copyright (c) 2026 Sustainable Global Societies Initiative

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
Abstract
Deciding whether a breast lesion is benign or malignant involves several image-reading steps, from mammography and ultrasound to histopathology, each open to disagreement between readers, yet most automated classifiers handle only one image type. We describe MORANet, a modality-oriented representation and attention network that keeps one architecture for all three imaging sources and changes only a short input-conditioning stage per modality. A pretrained EfficientNet-B3 encoder turns each image into feature tokens, multi-head self-attention lets the tokens exchange information across the image, a bidirectional GRU reads them in both directions, and a pooled dense head returns the decision. A label-smoothed focal objective counters class imbalance, and Grad-CAM++ maps accompany each prediction. With separate models and stratified 70/10/20 splits, MORANet reached held-out balanced accuracies of 95.75–98.84% on the four BreakHis magnifications, 99.18% on CBIS-DDSM and 89.96% on BUSI, missing 18 of 1,557 malignant test images in total, and each added component improved on a plain EfficientNet-B3 baseline.