Neoplastic and Non-Neoplastic Acute Intracerebral Hemorrhage in CT brain scans: Machine Learning-based Prediction using Radiomic Image Features

Zielsetzung Early differentiation of neoplastic and non-neoplastic intracerebral hemorrhage (ICH) can be difficult in initial radio-logical evaluation, especially for extensive ICHs1. The aim of this study was to evaluate the potential of a machine learning based prediction of etiology for acute ICHs based on quantitative radiomic image features extracted from initial noncontrast-enhanced computed tomography (NECT) brain scans.

Material und Methoden The analysis included NECT brain scans from 77 patients with acute ICH (n=50 non-neoplastic, n=27 neoplastic). Radiomic features including shape, histogram and texture markers were extracted from non- , wavelet- and log-sigma-filtered images using regions of interest of ICH and perihematomal edema (PHE). 6090 quantitative predic-tors were evaluated utilizing random forest algorithms with 5-fold model-external cross-validation. Model stability was assessed through comparative analysis of 10 randomly drawn cross-validation sets. Classifier performance was compared with predictions of two radiologists employing Matthews correlation coefficient (MCC).

Ergebnisse ROC AUC of the test sets for predicting neoplastic vs. non-neoplastic ICHs was 0.89 (95% CI [0.68; 0.99]; P<0.001), specificities and sensitivities reached > 80%. Compared to the radiologists’ predictions, the machine learning algo-rithm yielded equal or superior results for all evaluated metrics. MCC of the proposed algorithm at its optimal operat-ing point (0.69) was significantly higher than MCC of the radiologist readers (0.54); P =0.01.

Schlußfolgerungen Evaluating quantitative features of acute NECT images in a machine learning algorithm provided high discriminatory power in predicting non-neoplastic vs. neoplastic ICHs. Utilized in clinical routine, the proposed approach could improve patient care at low risk and costs.