Research On Automatic Intelligent Segmentation Algorithm Of Pelvic CT Images

Pelvic abnormalities are one of the threats to human health worldwide,causing millions of deaths each year.Although early diagnosis and treatment can greatly improve the chances of survival,it remains an important challenge.With the development of deep learning technology,computer-aided pelvic segmentation technology is a promising method for the diagnosis and treatment of pelvic abnormalities.However,when it comes to pelvic anomaly diagnosis,the lack of training data and the low segmentation accuracy have hindered the development of computer-aided pelvic segmentation.To address the lack of training data,we constructed a large pelvic computed tomography(CT)image dataset PCT14 K,which enables the training of deep convolutional neural networks(CNN)for pelvic anomaly diagnosis and treatment.Aiming at the problems that pelvic tissue and surrounding tissue are easily confused,pelvic tissue shapes are diverse,and pelvic tissue overlaps,this paper proposes the following two methods based on deep CNN to significantly improve the accuracy of pelvic-related tissue segmentation.CT Image Segmentation Algorithms and Object Detection Assisted Pelvic CT Image Segmentation Algorithms:(1)Pelvic CT image segmentation algorithm based on multifrequency information fusion.The main improvements of the algorithm are:First,this paper designs a pelvic CT image segmentation algorithm MFFPNet(Multi-Frequency-Fuse-Path-Net)based on multi-frequency information fusion and proposes the idea of multi-frequency information fusion.The information of domain and spatial domain alleviates the problem of insufficient performance of existing segmentation algorithms in the face of challenges such as confusing pelvic tissue and surrounding tissue,and diverse pelvic tissue shapes.Second,the low-level semantic information guidance path and the intermediate semantic information guidance path are proposed and combined with the high-level semantic information path;the ability of pelvic tissue edge segmentation is further improved.In addition,the inference cost of the low-level semantic information-guided path and the intermediate-level semantic informationguided path is zero when the segmentation algorithm reasoning.Third,the combination of the attention refinement module and the learnable upsampling operation further enhances the feature extraction ability and feature discrimination of the segmentation algorithm alleviates the edge smoothing problem caused by frequent upsampling of the segmentation framework and alleviates the shape to a certain extent.Disadvantages of difficulty inaccurate segmentation of diverse pelvic tissues.Finally,to verify the effectiveness and efficiency of the algorithm,this algorithm comprehensively evaluated the proposed method on the open-source pelvic segmentation dataset and achieved a result of 77.164% m Io U(Mean Intersection over Union).(2)Target detection-assisted pelvic CT image segmentation algorithm.To alleviate the problem of blurred segmentation boundaries and confusion of segmentation categories due to overlapping between pelvic tissues,this section further expands the idea,adopts the feature of multi-task learning and mutual supervision,and uses the positioning ability of the detection network to assist the segmentation of pelvic CT images.The specific innovations are as follows: First,two segmentation head structures,Seg Head-RFPP and Seg Head-CSP3,are designed to alleviate the problem of blurred boundary segmentation caused by overlapping between pelvic tissues.Second,based on the two segmented head structures,the basic target detection framework is introduced,and the category-bounding box prior is introduced to simultaneously locate and segment the pelvic tissue in the form of target detection-assisted segmentation,which further alleviates the problems caused by the pelvic tissue.Overlap leads to the problem of class confusion.This algorithm has been tested on the open-source pelvic CT image segmentation dataset and achieved 95.4% m AP50(Mean Average Precision),78.6% m Io U(Mean Intersection over Union),99.6% Acc(Accuracy)target detection and segmentation accuracy,and the visualization results prove the algorithm.It can effectively alleviate the problem of overlapping between pelvic tissues.Experiments show that the multi-frequency information fusion pelvic CT image segmentation algorithm proposed in this paper effectively solves the problems that pelvic tissue and surrounding tissue are easily confused and the pelvic tissue is diverse.The target detection-assisted pelvic CT image segmentation algorithm proposed in this paper effectively alleviates the problem of overlapping between pelvic tissues.