ASLNet: an explainable deep learning framework for glioma grading and survival prediction

Area of Science:

• Neuroimaging
• Oncology
• Artificial Intelligence

Background:

• Arterial Spin Labeling (ASL) MRI offers noninvasive quantitative perfusion imaging.
• Vascular heterogeneity in gliomas correlates with aggressiveness and prognosis.
• Deep learning (DL) may enhance ASL's predictive capabilities for tumor grade and survival.

Purpose of the Study:

• To develop and validate ASLNet, an interpretable 3D residual network for predicting histopathologic grade and overall survival (OS) in diffuse glioma patients using ASL MRI.
• To assess the clinical utility of ASL-based DL for glioma risk stratification.

Main Methods:

• Retrospective analysis of 471 diffuse glioma patients with ASL MRI data.
• Development of two 3D residual networks: one for WHO grade classification and another for OS prediction (incorporating clinical variables).
• Performance evaluation using AUC, macro-F1 score, accuracy, and recall; saliency maps generated via integrated gradients.

Main Results:

• The ASLNet grading model achieved an AUC of 0.79 and macro-F1 of 0.74.
• The OS prediction model achieved an AUC of 0.70 and macro-F1 of 0.73, with high recall for the long-survival class (0.94).
• Saliency analysis identified hyperperfused tumor cores and peritumoral regions as influential for prediction, aligning with glioma biology.

Conclusions:

• ASLNet demonstrates the feasibility of interpretable, perfusion-based DL for glioma grade and survival prediction using ASL MRI.
• ASL MRI contains clinically relevant prognostic information for gliomas.
• ASL-based DL shows potential as a noninvasive tool for glioma risk stratification, pending external validation.