Weakly supervised attention-based learning supports breast cancer detection on deep ultraviolet-excitation fluorescence images

Abstract Microscopy with ultraviolet surface excitation (MUSE) enables rapid fluorescence imaging of fresh tissue surfaces without conventional sectioning, but its appearance differs from that of haematoxylin and eosin-stained slides. Pixel- or patch-level annotation for deep learning is laborious, particularly for emerging imaging modalities. Here, we evaluated weakly supervised learning (WSL) for breast cancer detection on MUSE images using attention-based multiple instance learning trained with microscopic image-level labels alone. Fresh breast tissues from 35 mastectomy patients were stained with Hoechst 33,342 and terbium and imaged by MUSE. We analysed 700 images, comprising 10 cancerous and 10 non-cancerous images per case. Data were split at the case level into an independent test set of 140 images and a training/validation set of 560 images with fixed four-fold cross-validation. To assess the effect of patch size, we trained models using 224 × 224- or 64 × 64-pixel patches. The models achieved high test performance, with receiver operating characteristic area under the curve values of 0.983 and 0.979 for the 224 × 224- and 64 × 64-pixel models, respectively. Attention maps highlighted tumour-rich regions and cancer-associated stroma. These findings support the feasibility of WSL for MUSE-based breast cancer detection while reducing the need for exhaustive annotation.