Purpose: To investigate the feasibility of texture analysis of apparent diffusion coefficient (ADC) maps for differentiating fat-poor angiomyolipomas (fpAMLs) from non-clear-cell renal cell carcinomas (non-ccRCCs).
Methods: In this bi-institutional study, we included two consecutive cohorts from different institutions with pathologically confirmed solid renal masses: 67 patients (fpAML = 46; non-ccRCC = 21) for model development and 39 (fpAML = 24; non-ccRCC = 15) for validation. Patients underwent preoperative magnetic resonance imaging (MRI), including diffusion-weighted imaging. We extracted 45 texture features using a software with volumes of interest on ADC maps. Receiver operating characteristic curve analysis was performed to compare the diagnostic performance between the random forest (RF) model (derived from extracted texture features) and conventional subjective evaluation using computed tomography and MRI by radiologists.
Results: RF analysis revealed that grey-level zone length matrix long-zone high grey-level emphasis was the dominant texture feature for diagnosing fpAML. The area under the curve (AUC) of the RF model to distinguish fpAMLs from non-ccRCCs was not significantly different between the validation and development cohorts (p = .19). In the validation cohort, the AUC of the RF model was similar to that of board-certified radiologists (p = .46) and significantly higher than that of radiology residents (p = .03).
Conclusions: Texture analysis of ADC maps demonstrated similar diagnostic performance to that of board-certified radiologists for discriminating between fpAMLs and non-ccRCCs. Diagnostic performances in the development and validation cohorts were comparable despite using data from different imaging device manufacturers and institutions.
Keywords: Angiomyolipoma; Artificial Intelligence; Diffusion magnetic resonance imaging; Machine learning; Magnetic resonance imaging; Renal cell carcinoma.
Copyright © 2021. Published by Elsevier B.V.