Scribble-supervised medical image segmentation based on dynamically generated pseudo labels via multi-scale superpixels

Nowadays, deep learning-based training increasingly requires adequate pixel-level labeled data. However, in the context of medical image segmentation, accurate pixel-level labeling of lesion edges remains a significant challenge for annotators. Nevertheless, making medical image segmentation with weak annotations is one of the most difficult tasks currently because the weak annotations only cover a small portion of the image and contain little relevant information. To maximize the utility of weak annotations, we propose a novel segmentation method that relies on scribble annotations. By utilizing multi-scale superpixels and deep features from the U-Net, the proposed method iteratively expands and generates pseudo labels with higher accuracy and richer information. This process involves similarity calculations, a dynamic adjustment mechanism, and multi-scale refinement for epoch-wise network training. Thus, as the step-wise expansion of high-confident pseudo labels and the elimination of low-confident ones, the performance of the method can gradually approach some fully-supervised methods. Our method outperforms other weakly annotated segmentation methods on the ACDC and ISIC2018 datasets, as shown by extensive experiments. The results show the segmentation performance of the proposed network is superiorly increased by approximately 1.8%, 3.5% and 1.6% on IoU, CPA and Dice, respectively, and the 95% hausdorff distance (HD95) decreased by approximately 0.8. Furthermore, ablation experiments confirm the effectiveness of each component of our method.