Research On Transformer-based Deep Learning Medical Image Organ Segmentation Method

Medical image segmentation aims to separate different tissues and structures in medical images.Deep neural networks have emerged as the main approach for medical image segmentation,due to the inherent limitations of convolutional operations,although convolutional neural networks(CNNs)have become the consensus for various medical image segmentation tasks,they have limitations in extracting remote image features.Transformer has shown excellent performance in extracting remote image features,but it cannot capture lowlevel features.Therefore,combining CNN and Transformer can lead to better results.In this paper,two algorithmic models that fully combine CNN and Transformer features are proposed for Transformer and deep learning medical image segmentation algorithms,and experiments are conducted on abdominal multi-organ and liver image datasets to improve the accuracy of medical image segmentation.The main research includes:(1)A Transformer-style module called MCT is proposed,which combines multi-scale contextual information from convolution and null convolution,and fully includes rich context between input keys on the feature map by modifying the key value acquisition in the Transformer module.An encoder-decoder model called MCTNet is also proposed in this paper.A loss function specifically for the multi-organ segmentation task is also proposed.Experimental results show that the proposed MCTNet outperforms other competing methods in different evaluation metrics on medical image segmentation datasets.(2)A novel neural network model,using Transformer as the infrastructure,is proposed for the medical image segmentation task.The model uses an attention mechanism to capture the contextual information in medical images and uses position coding to represent the location information of pixels,thus overcoming the input size limitation and insufficient contextual information problems of CNN models and improving the parallelization capability of the model.Also,the model proposes a new dual-feature fusion module,which fully fuses the local features obtained by the CNN layer and the global features obtained by the Transformer layer,and improves the network segmentation accuracy.The experimental results show that the model outperforms existing models on various medical image segmentation tasks.(3)A deep neural network-based 3D tissue segmentation and determination system for pig body CT data was constructed.By performing CT scans on live pigs,the system can quickly and accurately segment and determine their corresponding weight percentages without slaughtering the pigs,which is the first collaborative corporate research using this method for breeding pigs in China.The innovation of this technology is to breed good pigs without slaughtering them.The application of this system can improve the success rate and efficiency of production and provide a more intelligent and integrated comprehensive experience for future CT image quality breeding pig screening of live pigs.