| Objective:Urolithiasis is a global disease with a high incidence and recurrence rate,and stone composition is closely related to the choice of treatment,surgery,and preventive measures.However,there is still a lack of a simple,fast,non-invasive method for the composition analysis of urinary stone in vivo that is easy to be promoted clinically.Artificial intelligence(AI)combined with imaging,using advanced computer algorithms to deeply mine the multidimensional features embedded in medical imaging data,has been widely used in various fields of radiology,showing great potential and broad application prospects in disease diagnosis,lesion segmentation,tumor grading and staging,and prognosis prediction.The aim of this study is to establish an AI model using plain scan images of the urinary tract,and investigate the predictive ability and clinical application value of the model for different components of stones in the body,in order to provide urologists with more accurate and comprehensive stone diagnosis information and further guide the selection of treatment,surgery and preventive measures for patients with urinary stones.Materials and Methods:1.60 cases of calculi whose compositions were determined by infrared spectrum analysis were retrospectively collected.Two specialists predicted the composition of calculi based on CT images,stone morphology,average CT value and personal experience.The accuracy,sensitivity,specificity,positive predictive value,and negative predictive value of their multicategorical predictions were calculated,respectively.2.Retrospective analysis of CT plain images of patients with urinary stones whose composition was determined by infrared spectroscopy in a single center,including 337 cases of calcium oxalate monohydrate(COM)stones and 170 cases of non-COM stones.The images were manually segmented and the image features were extracted,randomly divided into the training set and the testing set in the ratio of 7:3.The least absolute shrinkage and selection operation algorithm(LASSO)was used to construct the AI model,and then the classification training and testing were carried out.3.Retrospective analysis was conducted on 1,880 cases of urinary calculi whose components were determined by infrared spectrum in Center A and Center B,including 1448 cases of COM,176 cases of anhydrous uric acid(AUA),97 cases of carbonate apatite(CA),95 cases of magnesium ammonium phosphate hexahydrate(MAP),and 64 cases of ammonium urate(AAU).All the images were manually segmented and the radiomics features were extracted.Data from Center A were randomly divided into training set and internal validation set in a 7:3 ratio,while data from Center B was used as external testing set.The LASSO algorithm was used to construct seven dichotomous classification models(COM vs.non-COM,AUA vs.non-AUA,AAU vs.non-AAU,MAP vs.non-MAP,CA vs.non-CA,AUA vs.AAU,MAP vs.CA)and one five-classification model(COM vs.MAP vs.CA vs.AUA vs.AAU)for internal validation and external testing,respectively.The dichotomous model was assessed by the area under the curve(AUC)of the receiver operating characteristic curve(ROC),and the multi-classification model was assessed by accuracy,precision,recall and F1 score.4.Thirty cases with the same CT model and scanning parameters were randomly selected from the data set of center B.The 1 mm thin layer images were reconstructed into two groups of images with layer thickness of 2 mm and 5 mm at the CT post-processing workstation.All CT images were manually segmented by two radiologists independently and then the image features were extracted.Reproducibility of image features was calculated for each case using the consistency correlation coefficient(CCC),with CCC < 0.85 being poorly reproducible.The reproducibility of the image features under different layer thicknesses was compared and analyzed.Results:1.For 60 urinary stones of four components,the prediction accuracy was27.08 %(95% CI: 0.1528-0.4185)for urologist A and 31.67 %(95% CI: 0.2026-0.4496)for urologist B.2.In the single-center study,the accuracy,sensitivity,and specificity of the AI model for predicting COM and non-COM stones were 88.3%,90.1%,and 84.3%,respectively,with an AUC of 0.933 for the validation set.3.In the bicentric study,the AUC of COM vs.non-COM,AUA vs.non-AUA,AAU vs.non-AAU,MAP vs.non-MAP,and CA vs.non-CA groups were 0.753,0.785,0.804,0.763,and 0.710,respectively.For the two subclasses of infected stones(MAP vs.CA)and the two subclasses of uric acid stones(AUA vs.AAU),the AUCs were0.714 and 0.838 for the two groups.The accuracy of the five classification models(COM vs.MAP vs.CA vs.AUA vs.AAU)in predicting the five types of stones was69.1%,94.3%,92.8%,82.5%,and 95.0%,respectively,and the prediction accuracy of MAP,CA and AAU were all over 90%.The predicted COM had an accuracy of0.703,a recall of 0.952,and an F1 score of 0.809,whereas the recall of the other four stones was less than 0.1 and the F1 score was less than 0.2.4.The mean CCC of all groups under the same doctor and different layer thickness image condition was less than 0.85(P < 0.05).Among the stone features obtained by the segmentation of the lesions by Doctor A,the features of poor reproducibility in 1mm vs.2 mm,2 mm vs.5 mm,and 1 mm vs.5 mm group were 50.13%,79.91%and 82.38%,respectively.The larger the thickness gap between CT images,the worse the reproducibility of the imaging features.Under the same layer thickness,there were 201(12.74%)poor reproducibility features between two doctors in the 1mm group,582(36.88%)in the 2 mm group,and 674(42.71%)in the 5 mm group.The thicker the layer thickness,the worse the reproducibility of the imaging features.Conclusion:1.The accuracy of specialists in predicting the composition of urinary stones by combining the morphological characteristics of stones,CT values and personal experience in treatment is not high.2.The artificial intelligence model based on single-center urological CT images achieves composition prediction of calcium oxalate monohydrate in vivo with good diagnostic performance,which is a novel,simple and non-invasive method for stone analysis.3.The bicentric AI model based on plain scan images of the urinary tract can accurately distinguish calcium oxalate monohydrate,magnesium ammonium phosphate,carbonate apatite,anhydrous uric acid and ammonium urate stones in vivo,and accurately distinguish different subtypes of infected stones and uric acid stones,providing urologists with more accurate and comprehensive stone diagnosis information.However,the diagnostic performance of some of the models is unstable,and its stability needs to be further optimized by expanding the sample size of stones.4.Different parameters have a significant impact on the reproducibility of image features,which is directly related to the robustness of the model.The application of artificial intelligence model in multi-center clinical practice still faces many problems and challenges. |
PDF Full Text Request