Abstract:

Objective To investigate the diagnostic value of ultrasound radiomics for the histology of substantial lesions in transplanted kidney. Methods A total of 186 allograft patients who underwent renal puncture biopsy due to abnormal creatinine were selected and divided into the acute rejection (AR) group (135 cases) and the tubular necrosis (ATN) group (51 cases) according to the biopsy results. The biopsy results and ultrasonic data of the transplanted kidney were collected. The diagnosis was made by two physicians according to conventional ultrasound parameters. Radiomics was applied for ultrasonic image feature extraction. Independent sample t test was used for the initial selection of all the acquired omics feature data, and then the least absolute shrinkage and selection operator (LASSO) algorithm were used to select the best effective features from the selected features. Random forest, K-nearest neighbor method, Logistic regression and support vector machine classifier were used to establish the prediction model. All patients were assigned to the training cohort and the validation cohort according to the ratio of 7∶3, and a 5-fold cross-validation strategy was used to analyze the accuracy, sensitivity, specificity and ROC area under curve (AUC) of each histological model validation cohort. Results In the physician group, the sensitivity was 56.2%, specificity was 60.7%, and accuracy was 57.5% of the differential diagnosis of AR and ATN based on conventional ultrasound parameters. The image omics method was applied to extract 137 histological features from each image, and 6 meaningful features were retained after screening, including Shape2D-Flatness, FirstOrder-Min, Histo-Min, Histo-VoxelCount, Grad-Std and GLCM-CS. The AUCs of random forest, support vector machine, Logistic regression and K-nearest neighbor method were 0.931 (95%CI: 0.779-0.997), 0.762 (95%CI: 0.604-0.897), 0.721 (95%CI: 0.582-0.808) and 0.713 (95%CI: 0.508-0.796) respectively, in which the sensitivity, specificity and accuracy of the random forest model were 97.60%, 80.00% and 85.80%, showing the best comprehensive performance. Conclusion Ultrasonography can extract more ultrasound image features, and each group of models has better differential diagnostic value for histological classification of renal allograft parenchymal lesions, which is superior to conventional ultrasound methods.

Key words: ultrasonography, graft rejection, kidney transplantation, kidney tubular necrosis, acute, artificial intelligence, histology, radiomics