Artificial Intelligence In Diagnosis Of Renal Artery Stenosis

ObjectiveTo investigate the value of auto-classification of renal arterial spectrum based on convolutional neural networks in evaluating renal artery stenosis.Materials and MethodsA study group was composed by 126 patients with intrarenal arterial spectrum by color Doppler flow imaging and results of renal arteriography(DSA).A total of 346 intrarenal arterial spectrum images were included.The intrarenal arterial spectrum images were divided into three groups based on the gold standard diagnosis:202 images in group A(diameter reduction<50%),16 images in group B(diameter reduction≥50%or<70%),128 images in group C(diameter reduction>70%).After random allocation,there were training sets(158 images of group A and 117 images of group C)and testing sets(44 images of group A and 11 images of group C)for CNN model.Before the training of CNN model,all the images were preprocessed by following steps:extraction from the original Dicom format,normalization of the coordinate axis of spectrum,standardization of image size and binarization.We design,build and train the convolution neural network(CNN).Then we compare the classification results of CNN model with ultrasonic doctors who diagnose RAS by intrarenal arterial spectrum to evaluate the performance level of auto-classification.ResultsIn the classification experiment of testing set,28 out of 44 images in group A and 7 out of 11 images in group C were classified correctly by CNN model.Thus the classification of renal arterial spectrum image based on CNN model in diagnosis severe RAS(diameter reduction≥70%)has acceptable accuracy,with ACC 63.6%,SEN 63.6%,SPEC 63.6%,PPV 30.4%,NPV 87.5%。Ultrasonic doctors classified all the images.The result is 171 out of 202 images in group A and 70 out of 128 images in group C were classified correctly.Thus the classification of renal arterial spectrum image based on ultrasonic doctors in diagnosis severe RAS(diameter reduction ≥70%)has higher accuracy but lower sensitivity than CNN model,with ACC 73.0%,SEN 54.7%,SPEC 84.7%,PPV 69.3%,NPV 74.7%。ConclusionOur CNN model for intrarenal arterial spectrum image recognition and classification in the screening of severe RAS gets a passably result in the experiment,with higher sensitivity than ultrasonic doctors.It will be helpful for inexperienced physicians in the CDS screening of severe RAS.This study is a new attempt which provides the basis for the auxiliary diagnosis of renal artery stenosis in the future.