Medical image segmentation is closely related to the accuracy of clinical imaging practice and diagnosis. However, the scarcity of labeled data due to rare cases and limited manpower and resources of radiologists in clinical settings has led to a shortage of annotated data. As a result, semi-supervised models have become an attractive approach. In this paper, we use an uncertainty-based estimation model for semi-supervised learning to demonstrate the effectiveness of semisupervised learning in improving the possibilities of medical image diagnosis, specifically in predicting and segmenting the left ventricle in cardiac images. In this paper, We employ the UA-MT framework, which is based on uncertainty modeling, for semi-supervised learning. We analyze and compare the segmented images generated by supervised and semisupervised learning using four labeled and eight unlabeled cardiac MRI datasets from four patients. The segmentation accuracy is evaluated through metrics such as Dice coefficient, Jaccard index, average surface distance (ASD), and Hausdorff distance (HD) to assess the differences in accuracy. Result finds that the semi-supervised learning approach improves the segmentation results by 10.24% in terms of Dice coefficient compared to supervised learning, with a decrease of 4.0 in ASD and a decrease of 8.67 in 95HD. It achieves favorable quantitative evaluation in the assisted segmentation diagnosis of left ventricle MRI in the heart. The conclusion can draw that this method exhibits higher robustness in terms of image segmentation accuracy compared to supervised learning. Its results can provide a basis for the analysis and diagnosis of left ventricular MRI images of the heart.
|