Predicting IDH Mutation in Glioma Patients Using Deep Learning Algorithms with Conformal Prediction

Area of Science:

• Neuro-oncology
• Artificial Intelligence
• Medical Imaging

Background:

• Glioma classification by the World Health Organization emphasizes genetic markers like isocitrate dehydrogenase (IDH) mutations.
• Accurate prediction of IDH mutation status is crucial for glioma patient diagnosis and treatment planning.

Purpose of the Study:

• To develop and evaluate an uncertainty-aware deep learning model for predicting IDH mutations in glioma patients.
• To assess the impact of conformal prediction (CP) for uncertainty quantification (UQ) on model performance.

Main Methods:

• Developed and compared 3D convolutional neural network (CNN) models and a logistic regression (LR) ensemble classifier.
• Employed conformal prediction (CP) with a 0.01 nonconformity threshold for UQ.
• Evaluated models on UCSF and UPENN datasets using Area Under Precision-Recall Curve (AUPRC) and Area Under Receiver Operating Characteristic (AUROC).

Main Results:

• 3D CNN models achieved validation AUPRC of 0.8102 and external test AUPRC of 0.2139.
• The LR ensemble model achieved an external test AUPRC of 0.2583.
• CP integration resulted in an AUROC of 0.8592 and AUPRC of 0.5217 on the external test set, demonstrating improved performance and reliability.

Conclusions:

• Uncertainty-aware deep learning models, particularly when enhanced with conformal prediction, show significant promise for predicting IDH mutations in glioma.
• CP effectively quantifies uncertainty, leading to more reliable predictions and improved diagnostic accuracy in neuro-oncology.