Design of CKDNet and PMGFANet segmentation models for medical images considering accuracy improvement and data scarcity resolution

Medical image segmentation can raise the diagnostic efficiency of doctors and reduce their burden, which is of great significance. To raise the accuracy of medical image segmentation and solve the problem of scarce medical image data, corresponding segmentation algorithms are designed separately. Firstly, a cascaded knowledge dissemination network model is designed to raise the accuracy of medical image segmentation through the collaborative work of three task subnets: initial segmentation, lesion classification, and fine segmentation, combined with a feature fusion module. Secondly, a semi-supervised segmentation model based on pseudo label guided feature aggregation network is designed, innovatively introducing pseudo label guided feature enhancement module and multi-scale multi-stage feature aggregation module, optimizing the teacher-student network architecture, and fully utilizing unlabeled data to supplement training samples, in order to solve the problem of data scarcity. The findings denoted that the average multi-category accuracy and segmentation accuracy of the cascaded knowledge dissemination network model were 86.7 % and 99.4 %. On the molecular nucleus segmentation dataset, when the proportion of annotated data was 10 % and 50 %, the maximum intersection to union ratios of the semi supervised segmentation model were 0.859 and 0.867, respectively. It can be seen that the cascaded knowledge dissemination network model has a segmentation accuracy of over 99.0 % on medical images, which can effectively improve the segmentation accuracy of medical images. The semi supervised segmentation model also has good intersection to union ratio and Dice coefficient in low annotated data ratios, and can complete segmentation tasks in medical image data scarcity. The two medical image segmentation models designed in the research have good effectiveness. The contribution of the research is the design of a semi supervised model consisting of a cascaded knowledge dissemination network model and a pseudo label guided feature aggregation network, which improves the accuracy of medical image segmentation and solves the problem of data scarcity.