Distinguishing Functional from Non-functional Pituitary Macroadenomas with a Machine Learning Analysis

Pituitary adenomas are rare intracranial tumors that are often found incidentally in MR images. On the other hand, radiomics is a new field whose aim is converting images in mineable data; particularly, texture analysis is a postprocessing technique extracting quantitative parameters from the heterogeneity of pixel grey level. In this scenario, machine learning can be applied in order to classify these adenomas into functional and non-functional starting from features extracted through texture analysis on MRI images acquired through a protocol including a coronal T2-weighted Turbo Spin Echo sequence. The boosting of J48, a multinomial logistic regression and K nearest neighbour are implemented employing Knime analytics platform. Excluding J48 whose accuracy was 83.0%, multinomial logistic regression and K nearest neighbour achieved accuracies beyond 92.0% and the Area Under the Curve Receiving Characteristic Operator till 98.4%. Diagnosing correctly this delicate disease is crucial in order to achieve the best management as well as the most appropriate cure for patients. The novelty of this paper lies in proving the ability of the combination of radiomics and machine learning to pre-operatively predict tumoral behavior. Prior to this analysis it was believed that only blood tests or histopathological analysis could provide this information.