Yolov9 models for cancer detection in biopsy images: segmentation on pan-cancer data

This work examines the use of recent Yolo models on segmenting cancer biopsy images to highlight the tissue classes of interest. The experiments include recent YoloV9 and V8 architectures, trained on a large pan-cancer dataset, which contains examples for 19 distinct cancer cases. The models’ predictive capability was measured and comparatively studied, and detailed Panoptic Quality and precision metrics are discussed. The results obtained indicate the fine performance of recent Yolo models on cancer segmentation task, while all version 9 algorithms did surpass preceeding architectures. This study also examined the effect of the Yolo framework data augmentation methods for performance enhancement, by conducting experiments for two image sizes, 256 (original size) and 512 pixels.