Research And Application Of Semi-supervised Medical Image Segmentation Algorithm

In the clinical applications,it is very important for doctors to accurately segment organs or diseased areas from medical images.With the emergence of medical segmentation methods based on convolutional neural networks(CNN)or vision transformers,there have been relatively complete landing applications using artificial intelligence algorithms for auxiliary diagnosis and rapid treatment planning.However,many medical segmentation algorithms need a large quantity of annotated data,and in terms of data annotation,the annotation of image semantic segmentation is one of the most expensive tasks.In medical image segmentation tasks,this usually requires dense pixel/voxel labeling,and with the expertise of doctors,labeling is costly in both labor and time.In this case,the semi-supervised medical segmentation algorithm has become an effective solution,which can make full use of unlabeled data,and some algorithms are comparable to supervised learning algorithms,thus greatly reducing the cost of labeling.The main research contents and innovations of this paper include the following three points:(1)The semi-supervised semantic segmentation algorithm SemiMixTransformer-UNet based on multiple uncertainties is proposed,which combines the advantages of Transformer and CNN,and combines the methods of domain adaptation and uncertainty estimation to accurately segment medical images.Industry best results have been achieved on the ACDC cardiac auto diagnosis and Synapse abdominal multiorgan datasets.(2)Taking few labels and light weight as the starting point,a multi-layer Dropout lightweight convolutional image segmentation network UNetTiny is designed.In order to further improve the performance of the model under small-sample semi-supervised training and make the model more lightweight,this paper proposed a cross-knowledge distillation algorithm SemiMixTransformer-UNet-Distllation based on domain adaptation,which improves the accuracy of the model and reduced the model parameters and accelerated model inference process.(3)Based on the above research of the algorithm,this paper designs and implements a semi-supervised medical auxiliary diagnosis system,and realizes the acceleration of the TensorRT model.The algorithm model is transplanted to the edge device Jetson-Nano,which can perform real-time auxiliary diagnosis and realize the real algorithm.engineering landing.Based on the above methods,this paper conducts experimental analysis through the public automatic heart lesion segmentation dataset ACDC and the Synapse abdominal multi-organ dataset.The experimental results on ACDC show that the algorithm can achieve an average dice cofficient of 0.914 in 45% of the data,which is better than the supervised learning algorithm SwinUNet full data training results,and surpasses basically all current semi-supervised learning algorithms.We only using 7%of the training data to achieve an average dice cofficient of 0.885,reaching the industry leading level,and the results also outperform other semi-supervised learning algorithms in the Synapse dataset.