CT-based machine learning model to predict the Fuhrman nuclear grade of clear cell renal cell carcinoma

• Published in: Abdominal imaging Vol. 44; no. 7; pp. 2528 - 2534
• Main Authors: Lin, Fan, Cui, En-Ming, Lei, Yi, Luo, Liang-ping
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.07.2019; Springer Nature B.V
• Subjects: Artificial intelligence; Bladder; CARCINOMAS; Classifiers; Clear cell-type renal cell carcinoma; Computed tomography; COMPUTERIZED TOMOGRAPHY; DECISION TREE ANALYSIS; Decision trees; DIAGNOSIS; Diagnostic systems; Feature extraction; Gastroenterology; Hepatology; Imaging; Kidney cancer; KIDNEYS; LEARNING; Learning algorithms; Lesions; Machine learning; Mathematical models; Medical imaging; Medicine; Medicine & Public Health; PATIENTS; Quality; Radiology; RADIOLOGY AND NUCLEAR MEDICINE; Retroperitoneum; Tomography; Ureters; Texture analysis; Clear cell carcinoma; Fuhrman nuclear grade; Machine learning
• ISSN: 2366-004X; 2366-0058; 2366-0058
• DOI: 10.1007/s00261-019-01992-7

Purpose To predict the Fuhrman grade of clear cell renal cell carcinoma (ccRCC) with a machine learning classifier based on single- or three-phase computed tomography (CT) images. Materials and methods Patients with pathologically proven ccRCC from February 1, 2009 to September 31, 2018 who were not treated were retrospectively collected for machine learning-based analysis. The texture features were extracted and ranked from precontrast phase (PCP), corticomedullary phase (CMP), nephrographic phase (NP) and three-phase CT images, and open-source gradient boosting from the decision tree library of CatBoost was used to establish a machine learning classifier to differentiate low- from high-grade ccRCC. The performances of machine learning classifiers based on features from single- and three-phase CT images were compared with each other. Results A total of 231 patients with 232 pathologically proven ccRCC lesions were retrospectively collected. 35, 36, 41, and 22 Features were extracted and ranked from PCP, CMP, NP, and three-phase CT images, respectively. The machine learning model based on three-phase CT images [area under the ROC curve (AUC) = 0.87] achieved the best diagnostic performance for differentiating low- from high-grade ccRCC, followed by single-phase NP (AUC = 0.84), CMP (AUC = 0.80), and PCP images (AUC = 0.82). Conclusion Machine learning classifiers can be promising noninvasive techniques to differentiate low- and high-Fuhrman nuclear grade ccRCC, and classifiers based on three-phase CT images are superior to those based on features from each single phase.