Multi Scale and Multi-Source Fusion Framework for Breast Cancer Detection Using YOLOv8

Breast cancer detection from histopathological images requires accurate analysis of both fine-grained cellular structures and broader tissue level context. Most approaches rely on microscopic images which capture detailed cellular features but often lack global structural information while whole slide images (WSIs) provide comprehensive tissue context but are computationally expensive to process directly. Therefore, this work proposes a multi-sourced learning framework that integrate microscopic BreakHis images with contextual information derived from WSIs. In addition, a patch-based strategy is employed to extract informative regions from WSIs using a sliding window approach; then an intensity-based filtering is used to remove background regions and texture-based filtering is used to eliminate non-informative patches. Furthermore, a novel “on the fly” fusion mechanism is introduced where each BreakHis image will be dynamically combined with a randomly selected WSI patch using weighted blending thus enhancing contextual representation while preserving label integrity and avoiding additional memory overhead. The fused dataset will be constructed in a yolo compatible format allowing for efficient training without storing intermediate data in memory. A pre-trained YOLOv8 nano model will be fine-tuned on the generated dataset for tumor detection and classification. Experimental results demonstrate improved performance in terms of precision, recall, and mAP@50 alongside stable convergence behaviour. As such the proposed framework effectively captures multi-scale features while maintaining computational efficiency making it suitable for real-time and large-scale medical imaging applications.