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

Introduction: Arterial spin labeling (ASL) MRI provides noninvasive quantitative perfusion information and may capture vascular heterogeneity associated with glioma aggressiveness and prognosis. Deep learning (DL) approaches applied directly to ASL volumes may therefore support imaging-based prediction of tumor grade and survival. This study aimed to develop and validate ASLNet, an interpretable three-dimensional residual network framework trained on ASL MRI to predict histopathologic grade and overall survival (OS) in patients with diffuse glioma. Methods: This retrospective study included 471 patients with histologically confirmed diffuse glioma who underwent ASL MRI between 2015 and 2021. Original three-dimensional ASL volumes were used as model inputs, together with isotropically resampled, Gaussian-smoothed, and z-score-normalized versions. Two custom three-dimensional residual networks were trained: one model for WHO grade classification and a second FiLM-type intermediate-fusion model incorporating ASL imaging with demographic and clinical variables, including age, sex, and extent of resection, for OS classification (<12 months vs. ≥12 months). Model performance was evaluated on held-out test sets using area under the receiver operating characteristic curve (AUC), macro-F1 score, accuracy, and recall. Integrated gradients were used to generate saliency maps and identify influential perfusion regions. Results: The ASLNet grading model achieved an AUC of 0.79, macro-F1 score of 0.74, accuracy of 0.60, and recall of 0.73. The OS prediction model achieved an AUC of 0.70, macro-F1 score of 0.73, accuracy of 0.66, and recall of 0.94 for the long-survival class. Saliency analysis highlighted hyperperfused tumor cores as influential regions for grade prediction, while survival prediction also involved peritumoral regions, consistent with biologically plausible patterns of perfusion heterogeneity in gliomas. Discussion: ASLNet demonstrates the feasibility of interpretable, perfusion-based deep learning for glioma grade and survival prediction using ASL MRI. These findings suggest that ASL contains clinically relevant information related to tumor vascular characteristics and prognosis. Although further external validation is warranted, this approach supports the potential clinical value of ASL-based DL as a complementary noninvasive tool for glioma risk stratification.