Research On Multi-organ Segmentation Algorithms Of Abdominal CT Images Based On Deep Learning

Abdominal multi-organ segmentation is of essential importance for disease diagnosis,surgical simulation,image-guided interventions,and especially in radiotherapy treatment planning.In clinical practice,manual segmentation by experts is time-consuming,laborintensive,and subject to subjective factors.The accuracy and stability of the segmentation results are difficult to guarantee.Automatic segmentation technology can assist doctors in accurately locating and segmenting organs of patients’ abdomen in a shorter amount of time,facilitating more precise surgical,chemotherapy,and radiotherapy treatments.However,due to the diversity in morphological structure,size,and imaging quality between diffierent images,automatic segmentation of multiple organs in abdominal CT images remains a challenging task and has become a hotspot in medical image analysis research.This paper conducted in-depth research on the automatic segmentation of multiple abdominal organs in CT images using deep learning methods,aiming to address segmentation difficulties such as fuzzy boundaries between multiple abdominal organs or tissue regions,category imbalances,insufficient annotated data,and unannotated data underutilization.The main work is as follows:(1)Aiming at the challenges of difficult segmentation due to fuzzy boundaries etc,this article proposes a network model for abdominal multi-organ segmentation based on region enhancement and edge detection.This method constructs a U-shaped encoder-decoder network with an edge detection module,multi-scale extraction module,and 2D dual-stream adaptive module.The edge detection module learns edge representations using fine-grained and structural information,which are then fused with segmentation features to provide guiding constraints.The multi-scale extraction module combines information from different scales to better extract features of multiple organs with varying sizes.The 2D dual-stream adaptive module refines representative features by allocating weights to k parallel convolution kernels using input-dependent spatial and channel attention mechanisms,thereby enhancing regional features.Additionally,this article introduce a deep supervision mechanism to improve gradient flow,optimizing segmentation through a combination of multi-stage outputs.Experimental results on the FLARE2021 and MARSS datasets show that this method outperforms advanced methods such as UNet,COBRA,and nn UNet.(2)Aiming at the issue that the learning of the organ region features is affected by nonorgan regions and 2D segmentation networks ignore the spatial relationships among multiple organs in 3D CT data,this paper proposes a bridged encoder-decoder method for abdominal multi-organ segmentation.Based on the whole image input strategy and coarse-to-fine segmentation framework,this method designs a bridged encoder-decoder network named Bridge Net capable of fully exploring globally and locally prominent features with different depths,spaces,and channels for abdominal multi-organ segmentation.The Bridge Net consists of an encoder,a multiple-to-single fusion module,a 3D dual-stream adaptive module,and a decoder.In this paper,the classic Seg Net is used to construct the encoder and decoder.The multiple-to-single fusion module extracts multi-scale global features and multi-dimensional local features from the encoder using parallel pooling and convolutional operation branches.The 3D dual-stream adaptive module further mines salient features in the extracted features.Experiments conducted on the FLARE2021 and MARSS datasets demonstrate that this method outperforms advanced methods such as UNet,VNet,and nn UNet.(3)Aiming at the issue of difficulty in acquiring annotated medical image data and underutilization of a large amount of unannotated data containing significant information,this study proposes a self-training and selected re-training-based multi-organ segmentation method for abdominal organs.The method improves the segmentation performance of the model by supplementing important information from a large amount of unannotated data.Specifically,the method constructs a self-training framework based on a coarse-to-fine segmentation model using Bridge Net.Unannotated images are selected based on the stability score of pseudo-labels saved during training,and the segmentation model is retrained accordingly.During retraining,strong data augmentation and test time augmentation were designed on unannotated images to reduce overfitting noise labels and decouple similar predictions between teacher and student models.In addition,single-label based connected component labeling was used for postprocessing.Experiments on the FLARE2022 dataset show that the proposed method outperforms advanced methods such as SSA and Siamese.