Artificial Intelligence Method Based On CT Image And Liquid Biopsy To Predict CcRCC And Non-ccRCC Classification

Objective: As the most common pathological type of renal cell carcinoma(RCC),renal clear cell carcinoma is highly malignant,prone to metastasis,and has a poor prognosis.Therefore,accurate diagnosis and differential diagnosis of clear cell renal carcinoma and non-clear cell renal carcinoma before operation are of great clinical significance for treatment and prognosis.Methods: 1.KiTS 19 database was downloaded,which included the CT enhanced cortical phase image data of 300 cases of renal tumors,including 210 cases with labels and 90 cases without labels.Firstly,based on the deep learning method of renal tumor CT image information,U-Net and Attention U-Net network architecture algorithms are used to construct the semantic segmentation model of renal tumor based on CT image and deep learning.Secondly,209 cases of clear cell renal carcinoma and 53 cases of non-clear cell renal carcinoma(31 cases of renal papillary cell carcinoma and 22 cases of renal chromophobe cell carcinoma)in Ki TS19 data set were selected.On the basis of the segmentation architecture of Attention U-Net network,3D CNN network architecture algorithm was added.Then the classification and prediction models of clear cell renal carcinoma and non-clear cell renal carcinoma were constructed.2.The hematological characteristics,urine indicators and CT images of 306 patients with RCC were collected,and the classification prediction of clear cell renal carcinoma and non-clear cell renal carcinoma based on machine learning of liquid biopsy indicators was conducted,and the correlation between these key indicators of liquid biopsy and CT images was analyzed.A liq＿ccRCC prediction model was constructed to accurately distinguish clear cell renal carcinoma and non-clear cell renal carcinoma.Results: 1.The Attention U-Net network model and patchwise-crop training method have better segmentation accuracy,and the predicted value is closer to the real value.The area under the ROC curve(AUC)value of the classification prediction model for distinguishing clear cell renal carcinoma from non-clear cell renal carcinoma based on CT images was 0.8095,the sensitivity was 0.7143,the specificity was 0.6875,and the accuracy was 0.7089.In order to test its generalization ability,the enhanced CT image data of 271 patients with RCC were collected from the Second Hospital of Lanzhou University,including 229 cases of clear cell renal carcinoma and42 cases of non-clear cell renal carcinoma(17 cases of renal papillary cell carcinoma and 25 cases of renal chromophobe cell carcinoma).The classification prediction model was tested,and the AUC value was 0.7658,the accuracy was 0.7232,the sensitivity was 0.7424,and the specificity was 0.6190.2.The AUC value of liq＿ccRCC prediction model was 0.9728,the accuracy was0.9372,the sensitivity was 0.9456,and the specificity was 0.9097.In order to test its generalization ability,79 liquid biopsy indicators of RCC patients were collected from the First Hospital of Lanzhou University to test the reliability of this method.The test results showed that the AUC value was 0.8507,the accuracy was 0.8222,the sensitivity was 0.8229,and the specificity was 0.8000.Conclusion: 1.The deep learning method based on the CT image information of renal tumors has realized the automatic segmentation of kidney and renal tumors and the automatic classification of clear cell renal carcinoma and non-clear cell renal carcinoma,which provides a new technology and method for the intelligent auxiliary diagnosis of renal tumor images.2.Liquid biopsy-based machine learning model(liq＿ccRCC)can predict the classification of clear cell renal carcinoma and non-clear cell renal carcinoma,and this non-invasive classification method has potential clinical value.