Cross-Attention Patch Fusion for Few-Shot Colorectal Tissue Generation
Mansoor Hayat, Armaan Dhaliwal, Muhammad Muneeb Ud Din, Rahat Izhar, Muhammad Nadeem, Nouman Ahmad
Abstract
Labeled colorectal histopathology is scarce, especially for rare patterns, which limits robust training and validation. We target patch-level classes (e.g., normal epithelium, tumorous glands, stroma) and propose Cross-Attention Patch Fusion, a few-shot generator that synthesizes class-conditioned tissue patches from only a handful of expert-labeled seeds. A base patch attends to k references to find semantically matched local blocks; these blocks are fused in feature space and reweighted with channel-and-spatial attention to preserve gland boundaries, epithelial contours, stromal context, and texture transitions. On NCT-CRC-HE, our method reduces Fréchet Inception Distance (FID) by 4.7%, raises LPIPS by 22.0% (to 58.57), and improves few-shot classifier accuracy by 4.8 percentage points when trained with our synthetic data. Qualitative results show realistic gland architecture and plausible stromal textures. In sum, cross-attention plus patch-wise fusion yields more realistic and diverse colorectal tissue patches from few examples and supplements rather than replaces expert labels.Clinical relevance— By generating realistic, diverse colorectal tissue images from only a few examples, our model can help train and validate AI diagnostic tools without the need for extensive manual annotation. This may accelerate development of automated classifiers for rare histological patterns, supporting earlier and more accurate detection of colorectal lesions in clinical practice