Joint task semi-supervised semantic segmentation for TRUS image

Transrectal ultrasound (TRUS) is an important tool for prostate observation, playing a crucial role in medical processes such as multimodal image registration and surgical planning. However, TRUS images are often plagued by blurry boundary information and artifacts, making manual annotation a time-consuming and labor-intensive task, which results in insufficient pixel-labeled data for training neural networks. Therefore, this paper introduces a semi-supervised neural network based on joint tasks for TRUS image segmentation, which combines confidence information from target localization and semantic segmentation to generate pseudo-labels for collectively training unlabeled samples. Additionally, this method supports the generation of pseudo-labels using only location information for weakly supervised training of neural networks. Compared to other state-of-the-art models, our results show that this approach outperforms in segmentation performance with a smaller dataset, achieving an mean Dice similarity coefficient and mean Jaccard similarity coefficient of 92.8 % and 88.4 %, respectively. Furthermore, the experiments reveal that reducing the coupling of two decoders' parameters between the segmentation and detection tasks improves segmentation accuracy as dataset size decreases.

Automated summary

In this paper, a semi-supervised neural network based on joint tasks is introduced for TRUS image segmentation. By combining confidence information from target localization and semantic segmentation, pseudo-labels are generated for collectively training unlabeled samples. Experimental results show that this approach outperforms other state-of-the-art models in segmentation performance with a smaller dataset, and reducing the coupling between tasks improves accuracy.