Automatic Assessment Of Split Renal Function Based On Plain CT And Artificial Intelligence

Accurate assessment of split renal function in patients with hydronephrosis is crucial for treatment options and prognosis prediction.At present,SPECT(single photon emission computed tomography)is the gold standard for clinical evaluation of split renal function,which is expensive,time-consuming,and has radiation hazards.Compared with SPECT,plain CT(computed tomography)has the advantages of low cost,fast imaging speed and low radioactivity.Renal morphological features based on plain CT images are related to renal function,and are expected to replace SPECT in evaluating split renal function.However,current related studies are all based on manual segmentation of relevant kidney regions and calculation of morphological features such as kidney volume,which is time-consuming,laborious,and highly subjective,the deep image features related to renal function have not been deeply excavated,and the performance of the evaluation model constructed is not good.Therefore,in response to the above problems,this paper proposes an automatic split renal function assessment method based on plain CT and artificial intelligence.The main tasks are as follows:First,aiming at the problem of accurate segmentation of renal parenchyma and hydronephrosis in plain CT images,a new segmentation network-CBAM-ResV-Netis proposed.In kidney segmentation,the boundary between some tissues and the background is blurred.At the same time,it is necessary to take into account the largevolume renal parenchyma area and the small-volume hydronephrosis area with various shapes,which makes segmentation difficult.The proposed network uses V-Net as the backbone network,and introduces the CBAM attention module before the V-Net downsampling convolution,which enables the network to more effectively focus on important kidney areas in CT images,and suppress useless background areas;using the Respath module to replace the skip connection part in V-Net helps to coordinate the features learned from images of different scales and improve the performance of the segmentation network.Second,an automatic assessment model of split renal function was constructed.Based on the automatic segmentation of the kidney region,this paper extracts 3390 radiomics features including morphology,first-order,texture and advanced features,and uses feature repeatability analysis,feature redundancy analysis,and maximum correlation and minimum redundancy analysis and 8 common classifiers such as logistic regression,random forest,Lasso regression,and elastic net regression to calculate the radiomics score.Afterwards,combined with radiomics scores and clinical characteristics such as gender and age,an automatic split renal function assessment model for patients with hydronephrosis was constructed.A total of 281 cases of data from two centers were included in this paper,and were divided into training set(159 cases)and test set(122 cases)according to different centers.The results show that the CBAM-ResV-Net proposed in this paper has a Dice similarity coefficient of 0.835 on the test set,and its performance is better than 3D UNet,UNETR,Swin-Unet network on the renal parenchyma and hydronephrosis area.Ablation experiments verified the effectiveness of the introduced CBAM module and Respath module.The split renal function assessment model based on radiomics scores and clinical characteristics had good performance(AUC=0.883),and there was no significant difference between the split renal function assessment model based on manual kidney segmentation and automatic kidney segmentation(P>0.05),automatic segmentation is expected to replace manual segmentation for automatic assessment of split renal function,so as to reduce radiation damage during assessment and improve the efficiency of clinical diagnosis.