A Novel Integrated Graph-Based Self-Supervised Image Segmentation With Application to Medical X-Ray Images

Image segmentation stands as a pivotal challenge in the realm of computer vision. Although, in recent times, deep learning-based segmentation methods have emerged as front-runners in furnishing cutting-edge solutions, a notable drawback of these techniques lies in their heavy reliance on extensive training data, a resource-intensive endeavor both in terms of time and cost. In limited data scenarios such as medical and military applications, the associated challenges are even more pronounced. As such, there remains a pressing need for approaches that can mitigate this dependency on extensive data while still delivering robust and accurate segmentation results. An attractive alternative solution is self-supervised image segmentation techniques that bypass the need for large training data. So, this article uniquely formulated the image segmentation problem as a self-supervised graph clustering problem wherein different image descriptors were explored to discover the optimal graph representation for an image. The segmentation problem was then solved using a deep modularity network (DMoN), a recently proposed self-supervised graph clustering algorithm. To improve the segmentation performance further, the proposed DMoN-based model was fused with other unsupervised models like spectral clustering and the shortest path model to obtain a state-of-the-art integrated image segmentation model. To reflect the efficacy of the proposed integrated model, it was applied to the Shenzhen chest X-ray image database, wherein it outperformed the established techniques and achieved an impressive average Dice coefficient of 0.883.