Identification Of Rib Fracture Lesions Based On CT Images Of The Chest

Objective:Rib fracture,specifically a complete or partial break in the continuity of the rib structure,is one of the most common thoracic traumas in clinical practice.Chest computed tomography is a common method to diagnose rib fractures,however,due to the large number of rib fractures and their disorganized locations,doctors are prone to miss and misdiagnosis when reviewing the films,which may lead to doctor-patient disputes and even delayed treatment,so there is an urgent need for a method that can help doctors accurately and quickly detect specific rib fracture areas.Deep learning has been widely used in the field of medical image processing in recent years,and with its powerful sample learning and analysis capability,it is expected to further realize artificial intelligence-assisted medical diagnosis.Therefore,this paper applies deep learning algorithms around the recognition of rib fracture lesions based on CT images of the chest.Methods:Firstly,this paper optimizes the classical 3D image segmentation models 3D U-Net network and V-Net network respectively,changes their layer structure and replaces the original activation function with LeakyReLU,and proposes LUNet and LVNet respectively,making them more applicable to the segmentation of rib fracture region and improving the segmentation accuracy.Secondly,an improved 3D U-Net network(NOL-LUNet)based on a lightweight non-local block is proposed.Based on the LUNet network proposed in the previous paper,the non-local module is further optimised at the 3D level,and then the 3D non-local block is added to two specific downsampling layers of the LUNet network,which can effectively obtain global information,and the model is lightened by using group convolution at specific locations of the network to effectively reduce the amount of operations and the The number of parameters is reduced.This paper also proposes an improved V-Net network(CS-LVNet)based on the 3D CBAM attention mechanism.Based on the previously proposed LVNet network,the CBAM attention module is further optimised at the 3D level,and the 3D CBAM module is added to a specific downsampling layer of the LVNet network to further acquire spatial and channel information,which can effectively improve the accuracy for rib fracture segmentation.Results:In this paper,extensive experiments were conducted on RibFrac,an open dataset of rib fractures,and the proposed network model was experimentally proven to be very effective.First,compared with the original network,the PFs of the LUNet proposed in this paper were 1.47 lower and the Dice coefficient and IoU improved by 6.1%and 5.9%,respectively,while the PFs of the LVNet for detecting rib fractures were 14.334 lower and the Dice coefficient and IoU improved by 15.8%and 14.8%,respectively.Second,NOL-LUNet had a mean number of false positives per scan at maximum recall of 9.73,a Dice coefficient of 0.6547 and an IoU of 0.4866.Compared to using LUNet alone,the PFs were 8.84 lower and the Dice coefficient and IoU were 6.8%and 7.2%higher,respectively.Finally,the average number of false positives per scan for CS-LVNet at maximum recall was 10.8,with a Dice coefficient of 0.645 and IoU of 0.476.Compared to using LVNet alone,the PFs were 0.158 lower,and the Dice coefficient and IoU were 3%and 3.2%higher,respectively.All these results demonstrate the effectiveness of the proposed method for rib fracture lesion identification based on CT images of the chest.Conclusion:In this paper,a rib fracture lesion recognition algorithm based on CT images of the chest is investigated.After improving the classical 3D U-Net network and V-Net network,LUNet and LVNet are proposed,and based on them,an improved 3D U-Net network based on lightweight non-local modules(NOL-LUNet)and an improved V-Net network based on 3D CBAM attention mechanism(CS-LVNet)are proposed respectively.The NOL-LUNet network is dedicated to the acquisition of global information,which further enhances the segmentation accuracy of the LUNet network and can effectively segment the rib fracture region.And the CS-LVNet network further improves the accuracy of rib fracture lesion identification by embedding a 3D attention mechanism.Both methods are capable of rapidly extracting features of complex rib fractures,making the output intuitive,accurate and convenient,and have important academic and clinical implications for alleviating the over-reliance on CT image quality and physician experience in the diagnosis and treatment of rib fractures in hospitals and judicial institutions.