Identification of sarcomatoid feature and nuclear grades of clear cell renal cell carcinoma based on artificial intelligence network of computerized tomography

林博涵,邱钱仁顺,陈少豪,许宁,郑清水,薛学义,魏勇
DOI: https://doi.org/10.3760/cma.j.cn421213-20210907-01262
2022-01-01
Abstract:Objective:To construct an artificial intelligence (AI) network based on preoperative computerized tomography (CT) to identify sarcomatoid feature and nuclear grades of clear cell renal cell carcinoma (ccRCC).Methods:The preoperative CT imaging and clinicopathological data of 286 patients with ccRCC from April 2015 to August 2018 were analyzed retrospectively. The AI network model based on preoperative CT was established and the role of the model in identifying high and low nuclear grades and sarcomatoid differentiation of ccRCC was tested. The t-test, receiver operating characteristic (ROC) curve and Delong’test were used for statistical analysis. Results:Among the all patients, 22 cases (7.7%) were accompanied by sarcomatoid differentiation. The pathological nuclear grade was grade 1, 39 (13.6%); grade 2, 186 (65.0%); grade 3, 43 (15.0%); and grade 4, 18 (6.4%). AI predicted the probability of sarcomatoid feature in sarcomatoid differentiation group was significantly higher than that in non-sarcomatoid differentiation group [(73.66±21.99)% vs. (37.41±23.40)%, t=-7.012, P<0.01], area under the ROC curve (AUC)=0.87 (0.80-0.94). There was no significant difference in the maximum tumor diameter between them ( t=-0.538, P>0.05). The maximum tumor diameter in the low-grade group was significantly smaller than that in the high-grade group [(4.83±3.10) cm vs. (5.85±3.69) cm, t=-2.166, P<0.05]. The probability of AI predicting the low-grade pathology in the low-grade group was significantly higher than that in the high-grade pathology group [(74.96±18.52)% vs. (40.28±0.21)%, t=12.713, P<0.01], AUC=0.88 (0.83-0.93), which performed better than that in the maximum tumor diameter of 0.62 (0.54-0.70) (Delong’test, Z=4.563, P<0.01). Conclusion:The AI network based on preoperative CT is helpful in identifying the sarcomatoid feature and nuclear grades of ccRCC.
What problem does this paper attempt to address?