Specific Instance and Cross-Prompt-Based Robust 3-D Semi-Supervised Medical Image Segmentation

In clinical applications, accurate segmentation of lesions or organs is crucial to the subsequent disease diagnosis. Due to the expensive and labor-intensive nature of annotations, semi-supervised learning (SSL) becomes a practical solution to the shortage of labeled images. Among the various SSL models, mean teacher (MT) has demonstrated encouraging performance in the segmentation of 3-D medical images. Nevertheless, the unlabeled data is treated equally and its discrepancy is generally ignored in the model, which significantly affects the segmentation performance. Besides, the scarcity of 3-D medical samples and lack of guidance in the image segmentation further aggravate the degradation of performance. In order to tackle the above issues, a specific instance and cross-prompt-based model is presented for 3-D medical image segmentation. Specifically, we design an instance-specific adaptive module including different image filters to adaptively enhance each instance upon the propensity score. In such a way, the model gradually absorbs the external instructive knowledge from each specific instance according to the gap between the student and teacher models, which will improve the model's ability for generalization. Moreover, the inevitable conflicts between the prediction of the teacher model and ground truth (GT) will limit the effectiveness of the student model, the cross-prompt strategy combined with a guided filter network is introduced to enhance the segmentation performance by refining the segmentation map for preserving the image edge and details while alleviating the coupling between student and teacher. Three public datasets and one in-house dataset are adopted to validate the superiority and robustness of our algorithm. The proposed approach can not only achieve remarkable segmentation results but also outperforms other mainstream SSL segmentation algorithms in both qualitative and quantitative evaluations. The code will be accessed at http://github.com/Fyw1988/MUL_SICP .