Aim: To evaluate the automated medical decision support system "Sechenov.AI_nephro" in the treatment of patients with renal parenchymal tumors.
Materials and methods: The beta version of the web-platform "Sechenov.AI_nephro" consists of a neural network based on MONAI (Medical open network for AI) and a web interface, with algorithms classified based on segmentation data in manual mode using the 3D modeling program "Amira". A total of 441 patients with renal parenchymal tumors were included in the multicenter prospective study. Testing was carried out over 12 months in 3 urological centers: 358 (81.2%) patients from I.M. Sechenov First Moscow State Medical University, Moscow; 73 (16.6%) patients from Bashkir State Medical University; and 10 (2.3%) patients from Saratov State Medical University named after V.I. Razumovsky. In all cases, contrast-enhanced computed tomography (CT) was performed preoperatively. DICOM (Digital Imaging and Communications in Medicine) data of each patient's CT was uploaded to the web-platform "Sechenov.AI_nephro" for automatic construction of a 3D model of the tumor. The work of the web-platform "Sechenov.AI_nephro" was evaluated based on a questionnaire completed by surgeons who performed the surgical intervention. The questionnaire consisted of 14 questions, with a scoring system from 1 to 10 points. It was divided into 3 main sections, including first for assessment of the quality of work of the web-platform "Sechenov.AI_nephro"; second for evaluation of the use of the 3D model in communication with the patient, for surgical planning and intraoperative navigation; and third for analysis of the choice of useful data display mode, errors in constructing the 3D model.
Results: The questionnaire was completed in 253 (57.37% of 441) cases. The quality of 3D models was rated 7.8-9.4 points, and the use of the 3D model in communication with the patient, for surgical planning and intraoperative navigation was rated 7.8-9.4 points. The 3D models were constructed correctly in 70% of cases. The area of interest was the useful mode of 3D models display in surgical planning. Incorrectly constructed anatomical elements were veins in 25.5% and the tumor in 26.4% of cases, respectively.
Conclusion: The automated medical decision support system in the treatment of patients with renal parenchymal tumors "Sechenov.AI_nephro" demonstrated satisfactory quality of 3D reconstruction of pathological process. 3D models allow for personalized determination of the surgical tactic for treating patients with renal tumors.
Keywords: 3D modeling; artificial intelligence; computer vision; kidney cancer; machine learning; partial nephrectomy.