DRCL: rethinking jigsaw puzzles for unsupervised medical image segmentation

Data that are manually annotated is essential for the training of deep neural networks. However, in the field of medical image segmentation, the annotation process is often prohibitively expensive and inefficient. Unsupervised medical image segmentation offers a promising alternative but faces challenges in feature representation and utilization. To address these issues, we propose the dual reconstructed contrastive learning (DRCL) framework, specifically designed for unsupervised medical image segmentation. This framework incorporates a dual reconstructed contrastive learning strategy, analogous to the assembly of a jigsaw puzzle, which reconstructs the original image during local contrastive learning, thereby facilitating the generation of a latent space tailored to the segmentation task. Additionally, the integration of spatial coordinates ensures the preservation of spatial relationships throughout the transformation procedure. Subsequently, deep spectral clustering is applied for a secondary analysis of the latent space features, thereby augmenting their utilization. Experiments conducted on three datasets-encompassing 56, 100, and 200 samples of retinal atrophy, brain ventricles, and brain tumors, respectively-reveal that the DRCL framework significantly outperforms state-of-the-art unsupervised segmentation methods, with respective improvements in Dice scores of 8.2%, 3%, and 6.4%. Notably, the DRCL framework achieves performance comparable to the segment anything model (SAM) on the retinal atrophy dataset and exceeds SAM's performance on the other datasets, despite SAM's extensive pre-training on 11 million images and 1.1 billion masks. These findings highlight the superior effectiveness of our proposed method compared to existing approaches. The source code is accessible at https://github.com/wwzz-max/DRCL.