Aim: To evaluate the possibilities of textural analysis of 3D models in differentiating the degree of nuclear dysplasia of the clear cell renal cell carcinoma (ccRCC).
Materials and methods: The specimens after surgical treatment of 190 patients with ccRCC were analyzed. In all cases, nephron-sparing surgery (NSS) was performed through laparoscopic access. The clinical characteristics were evaluated, including age, gender, tumor localization (side, surface and segments), absolute tumor volume, Charlson comorbidity index, body mass index, nephrometry scores (RENAL, PADOVA, C-index). Patients were divided into 2 groups. In group 1, there were 119 patients with the ccRCC of Grade 1 or 2, while group 2 consisted of 71 patients with ccRCC of Grade 3 and 4. All patients underwent 3D virtual planning of procedure using the 3D modeling program "Amira". At the first stage, two experienced radiologists performed manual segmentation of 3D models of kidney parenchyma tumors. At the second stage, the tumor shape was analyzed with a mathematical calculation of three indicators and more than 300 textural features of statistics of types 1-2 were extracted. Further, an intellectual analysis was carried out. For the evaluation of tumor grade according to Furman system, the classification problem was solved using the machine learning algorithm Stochastic Gradient Descent and cross-validation k=5.
Results: The accuracy of classification for the two groups of Grade 1 or 2 and Grade 3 or 4 on the F1 metric was 72.2. To build the model, the following parameters were selected: the absolute tumor volume, the Charlson comorbidity index, "Energy", the first quartile and the second decile of the pixel intensity distribution.
Conclusion: The texture analysis of 3D models for the prediction of Fuhrman grade in ccRCC demonstrated satisfactory quality for two groups of Grade 1 or 2 and Grade 3 or 4 nuclear dysplasia.
Keywords: 3D technologies; kidney parenchyma cancer; machine learning; texture analysis.