Unconditionally Generated and Pseudo-Labeled Synthetic Images for Diabetic Foot Ulcer Segmentation Dataset Extension

Abstract

The diabetic foot syndrome is a long-term complication of diabetes mellitus. Affected persons are prone to acquisition of deep tissue injuries due to neuropathy-related sensory impairment. When not detected early, Diabetic Foot Ulcers (DFUs) can manifest. Impaired healing capabilities, given vascular damage, and wound bed colonization promote chronification. Subtle changes in the wound bed indicate complications demanding intervention, otherwise prolonging need for treatment. Hence, to enable proper healing frequent and detailed monitoring is necessary. Deep learning-based segmentation is a key technology for automated DFU analysis at the point-of-care, enabling fast measurement and comprehension of subtle changes over time. The research at hand investigates an approach on semantic DFU segmentation, developed during participation in the DFU Challenge (DFUC) 2022. It involves a large ensemble of 25 models of a Feature Pyramid Network with an SE-ResNeXt101-32x4d backbone. Models were trained on an extended training set, enriched to three times its original size via pseudo-labeled synthetic images, generated via the data-efficient unconditional StyleGAN2+ADA. Segmentation performance achieved by challenge submissions is reported. Further, results of synthetic image generation are presented, achieving notably good quality. Results show that the approach achieved competing results, yet overfitting to the synthetic extension was observed. A critical discussion addresses method potentials and risks, points out limitations, and suggests improvements. The work concludes that training set extension with unconditionally generated and pseudo-labeled synthetic images can be achieved rather effortlessly, but increases computational costs and experiment complexity.