Explainable diffusion-perfusion radiomics using apparent diffusion coefficient and arterial spin labeling for ATRX status prediction in glioblastoma.
Authors
Affiliations (9)
Affiliations (9)
- Division of Neuroimaging and Neurointervention, Department of Radiology, Stanford University, Stanford, CA, United States.
- Department of Information Technology, K. J. Somaiya School of Engineering, Somaiya Vidyavihar University, Mumbai, India.
- Faculty of Medicine, Ovidius University of Constanta, Constanta, Romania.
- Department of Electrical and Computer Engineering, National Technical University of Athens NTUA, Athens, Greece.
- Department of Medicine, National and Kapodistrian University of Athens, Athens, Greece.
- Department of Biological Sciences, Kean University, Union, NJ, United States.
- Internal Medicine-Hematology, University of Patras Medical School, Patras, Greece.
- 1st Department of Neurology, Medical School, National and Kapodistrian University of Athens, Eginition Hospital, Athens, Greece.
- Department of Radiology, University of Miami, Miami, FL, United States.
Abstract
ATRX status serves as a key biomarker for the molecular profiling of glioblastoma, but its evaluation currently depends on invasive tissue sampling. While radiomics has demonstrated promising potential in the molecular analysis of gliomas, the utility of physiological MRI biomarkers for ATRX prediction remains underexplored. Our goal was to develop and assess an explainable combined model utilizing the apparent diffusion coefficient (ADC) and arterial spin labeling (ASL) for the noninvasive prediction of ATRX. We retrospectively included 106 patients with histologically confirmed glioblastoma from the UCSD-PTGBM cohort. Radiomics features were extracted from ADC and ASL imaging sequences and used to create a combined machine learning model. The model's performance was assessed using metrics including the area under the receiver operating characteristic curve (AUC), precision-recall analysis, balanced accuracy, calibration metrics, and decision curve analysis. SHapley Additive exPlanations (SHAP) were employed to identify the imaging features most influential to model predictions. Our integrated ADC and ASL model achieved an AUC of 0.711, with sensitivity 0.737, balanced accuracy 0.713, specificity 0.690, and NPV 0.923 (driven in part by the low prevalence of ATRX loss in the cohort). SHAP analysis revealed that ADC- and ASL-derived features accounted for 48.5% and 51.5% of total mean absolute SHAP importance, respectively, indicating that features from both modalities contributed to predictions within the jointly fitted model, while their independent incremental value remains to be established. Decision curve analysis showed a net benefit at threshold probabilities of roughly 0.01 to 0.46. These findings suggest that combining diffusion and perfusion radiomics can provide noninvasive insight into ATRX status in glioblastoma, though the model's moderate calibration and modest positive predictive value indicate it is best regarded as a hypothesis-generating screening adjunct rather than a diagnostic tool. Radiomics-based machine learning models from physiological MRI may offer valuable insights into ATRX status in glioblastoma; further validation with larger, independent patient cohorts is warranted.