3D MRI-Based Topological Analysis with Machine Learning to Predict Skull Base Meningioma Pathologic Grade

Introduction: WHO pathologic grade serves as a key element in driving clinical management of skull base meningiomas, with grade 2 and 3 lesions requiring different surgical, radiation, and surveillance strategies compared to grade 1 lesions. We currently have limited ability to predict meningioma grading preoperatively. Here, we apply machine learning to a 3D, MRI-based analysis of tumors’ topologic and geometric features to predict pathologic grade of skull base meningiomas using imaging alone.

Methods: We analyzed preoperative MRIs from 38 patients with skull-base meningiomas. Each MRI was segmented into three separate masks: enhancing tumor, non-enhancing/necrotic tumor, and surrounding FLAIR abnormality. The segmentation masks were processed to create topological features (chosen a priori) describing the tumors’ 3D structures. These features, without other clinical variables, were used in a custom machine learning pipeline to predict binarized pathologic grade of the tumor (grade 1 vs. grade 2/3). The machine learning algorithm was XGBoost, owing to its traditionally strong performance on tabular data. Only one algorithm was tested due to the already-small sample size and the possible model instability that could arise from additional data splitting. We split data using a nested cross-validation scheme and evaluated model performance with area under the receiver operating characteristic curve (AUROC). Performance was also assessed against a random classifier.

Results: In sum, there were 31 patients with grade 1 tumors, five with grade 2, and two with grade 3. The 3D segmentations were processed to create rotation-invariant geometric features (shape histograms, D2 histograms, shape PCA histograms) and topological features (connected components) ([Fig. 1]). The final AUROC for the machine learning model was 0.75 (bootstrapped 95% confidence interval, 0.46 to 0.97; [Fig. 2]). Though this confidence interval is wide, when compared to the random classifier (with expected AUROC 0.50), the custom algorithm performed significantly better (p = 0.045).

Conclusions: Based on preliminary analysis of this small patient sample, topological analysis of preoperative skull-base meningioma MRIs may enable prediction of WHO pathologic grade. After more rigorous study with larger sample size and prospective validation, tools such as this may enable improved prognostication for this patient population and aid in counseling patients.

Fig. 1 Segmentation masks and examples of geometric/topological features.

Fig. 2 Receiver operating characteristic (ROC) curve of machine learning model on final test set data.

Publikationsverlauf

Artikel online veröffentlicht:
05. Februar 2024