Abstract:
Segmentation of pancreatic cancer in CT images using supervised algorithms, despite technological progress, still necessitates extensive manual annotations (segmentation masks). These annotations are costly and may introduce biases. This research introduces a novel approach as weak supervision anomaly detection for pancreatic tumor detection, employing denoising diffusion algorithm, only using healthy/diseased labels. Through a deterministic iterative noise manipulation process and classifier guidance, it enables seamless image translation between healthy and diseased subjects, producing detailed anomaly maps without the need for complex training or segmentation masks.