Semi-Supervised Medical Image Segmentation using Puzzlemix Augmentation Technique

The task of fully supervised segmentation in medical image analysis is challenging, requiring a substantial volume of manually annotated data. Unfortunately, many existing learning-based approaches face challenges due to the scarcity of such annotated medical data. This scarcity poses a significant practical obstacle to achieving accurate and robust medical image segmentation. In this work, we present a simple ’Mean Teacher’ architecture by using cut-mix and puzzle mix augmentation technique between limited scribbled data and unlabeled data. This simple approach works surprisingly well, performing on the comparable level as previous fully supervised methods with scribbled labeled data. Evaluation of these techniques is conducted on publicly available cardiac (ACDC) and MSCMR segmentation datasets.