Generating Progressive Images from Pathological Transitions Via Diffusion Model

Zeyu Liu, Tianyi Zhang, Yufang He, Guanglei Zhang

Abstract: Pathological image analysis is a crucial field in deep learning applications. However, training effective models demands large-scale annotated data, which faces challenges due to sampling and annotation scarcity. The rapid developing generative models show potential to generate more training samples in recent studies. However, they also struggle with generalization diversity when limited training data is available, making them incapable of generating effective samples. Inspired by pathological transitions between different stages, we propose an adaptive depth-controlled diffusion (ADD) network for effective data augmentation. This novel approach is rooted in domain migration, where a hybrid attention strategy blends local and global attention priorities. With feature measuring, the adaptive depth-controlled strategy guides the bidirectional diffusion. It simulates pathological feature transition and maintains locational similarity. Based on a tiny training set (samples ≤ 500), ADD yields cross-domain progressive images with corresponding soft labels. Experiments on two datasets suggest significant improvements in generation diversity, and the effectiveness of the generated progressive samples is highlighted in downstream classification tasks.