A Novel Hierarchical Generative Model for Semi-Supervised Semantic Segmentation of Biomedical Images

In biomedical vision research, a significant challenge is the limited availability of pixel-wise labeled data. Data augmentation has been identified as a solution to this issue through generating labeled dummy data. While enhancing model efficacy, semi-supervised learning methodologies have emerged as a promising alternative that allows models to train on a mix of limited labeled and larger unlabeled data sets, potentially marking a significant advancement in biomedical vision research. Drawing from the semi-supervised learning strategy, in this paper, a novel medical image segmentation model is presented that features a hierarchical architecture with an attention mechanism. This model disentangles the synthesis process of biomedical images by employing a tail two-branch generator for semantic mask synthesis, thereby excelling in handling medical images with imbalanced class characteristics. During inference, the k-means clustering algorithm processes feature maps from the generator by using the clustering outcome as the segmentation mask. Experimental results show that this approach preserves biomedical image details more accurately than synthesized semantic masks. Experiments on various datasets, including those for vestibular schwannoma, kidney, and skin cancer, demonstrate the proposed method's superiority over other generative-adversarial-network-based and semi-supervised segmentation methods in both distribution fitting and semantic segmentation performance.