Machine learning-based texture analysis for differentiation of large adrenal cortical tumours on CT

AIM: To compare the efficacy of computed tomography (CT) texture analysis and conventional evaluation by radiologists for differentiation between large adrenal adenomas and carcinomas. MATERIALS AND METHODS: Quantitative CT texture analysis was used to evaluate 54 histopathologically proven adrenal masses (mean size=5.9 cm; range=4.1–10 cm) from 54 patients referred to Anderson Cancer Center from January 2002 through April 2014. The patient group included 32 women (mean age at mass evaluation=59 years) and 22 men (mean age at mass evaluation=61 years). Adrenal lesions seen on precontrast and venous-phase CT images were labelled by three different readers, and the labels were used to generate intensity- and geometry-based textural features. The textural features and the attenuation values were considered as input values for a random forest-based classifier. Similarly, the adrenal lesions were classified by two different radiologists based on morphological criteria. Prediction accuracy and interobserver agreement were compared. RESULTS: The textural predictive model achieved a mean accuracy of 82%, whereas the mean accuracy for the radiologists was 68.5% (p<0.0001). The interobserver agreements between the predictive model and radiologists 1 and 2 were 0.44 (p<0.0005; 95% confidence interval [CI]: 0.25–0.62) and 0.47 (p<0.0005; 95% CI: 0.28–0.66), respectively. The Dice similarity coefficient between the readers' image labels was 0.875±0.04. CONCLUSION: CT texture analysis of large adrenal adenomas and carcinomas is likely to improve CT evaluation of adrenal cortical tumours.