E2137. Can Machine Learning of MRI Textural Features Differentiate Between Intra and Extra-Axial Brain Tumors?
Authors
Ohoud Alaslani;
The Ottawa Hospital
Nima Omid-Fard;
The Ottawa Hospital
Rebecca Thornhill;
The Ottawa Hospital
Nick James;
The Ottawa Hospital
Rafael Glikstein;
The Ottawa Hospital
Objective:
Determining the origin of intracranial lesions can be challenging in certain cases. To determine the feasibility of a machine-learning (ML) model in discriminating intraaxial (IA) from extraaxial (EA) lesions.
Materials and Methods:
Retrospective review of consecutive adult patients (> 18 years old) with a newly diagnosed solitary brain lesion, who underwent brain MRI at our institution from January 2017 to December 2018. Tumor volumes of interest (VOI) were manually segmented on both T2W and T1W postcontrast MR images. A machine learning algorithm, XGBoost, was used to generate classification models trained on the textural features extracted from the segmented VOIs, using histopathology as the reference standard.
Results:
A total of 92 lesions were analyzed, including 70 of IA and 22 of EA origin. Area under the ROC curve for the identification of intraaxial tumors was 0.91 (CI 0.89 - 0.93; Fig.1A) for the model based on T1-weighted postcontrast MRI features, 0.81 (CI 0.78 - 0.84; Fig.2A) based on T2-weighted MRI features, and 0.92 (CI 0.90 - 0.94; Fig.3A) based on all features. All models had high sensitivity in identifying IA tumors (> 90% for each), but none achieved high specificity (39 - 64%). Nevertheless, the models attained acceptable levels of accuracy (> 80%) and precision (> 88%).
Conclusion:
A location-based ML classification model for differentiating IA from EA tumors is feasible based on this preliminary study, with good sensitivity. However, specificity was low to moderate, likely due to the imbalanced dataset. Further study with balanced data and external validation is needed. ML algorithms may potentially aid in lesion localization when the origin is clinically ambiguous.