Comparative study of radiomics features between ASL and DSC perfusion-weighted imaging in distinguishing radiation-induced brain injury from glioma recurrence.

Authors

Abstract

To compare the radiomics features of pseudocontinuous arterial spin labeling (ASL) and dynamic susceptibility contrast (DSC) perfusion-weighted imaging in distinguishing radiation-induced brain injury (RIBI) from tumor recurrence (TR) in patients with diffuse gliomas. A total of 131 patients with pathologically confirmed adult-type diffuse gliomas, presenting with newly developed abnormal enhancing lesions after standard radiotherapy, were included for model development. Lesion segmentation was performed manually by two raters using ITK-SNAP and the segmentation consistency was assessed by intraclass correlation coefficient (ICC). A total of 1015 radiomics features were extracted from the segmented areas of T2WI, T1WI, contrast-enhanced T1WI (CE-T1WI), cerebral blood flow (CBF) maps from ASL, and relative cerebral blood flow (rCBF) maps from DSC, respectively. Feature selection was conducted via t-tests, and an elastic net regression (ENR) algorithm combined with recursive feature elimination (RFE) with cross-validation was used to identify the important features. Four machine learning classifiers were employed to establish radiomics models. Diagnostic performance was assessed using the area under the curve (AUC) in conventional (T2WI+T1WI + CE-T1WI), ASL-CBF, DSC-rCBF, conventional + ASL-CBF, and conventional + DSC-rCBF radiomics models, respectively. The optimal models were further validated in an independent validation set (n = 37). Six features from ASL-CBF, five from DSC-rCBF, one from T1WI, one from T2WI, and four from CE-T1WI were identified as the most significant features. The K-nearest neighbor (KNN) classifier was selected as the optimal classifier. In the training set, both ASL-CBF (AUC = 0.962) and DSC-rCBF (AUC = 0.946) radiomics models demonstrated satisfactory diagnostic performance, surpassing that of conventional MRI radiomics model (AUC = 0.844). The diagnostic accuracy was further improved when integrating DSC-rCBF features into conventional MRI models (AUC = 0.971). In the independent validation set, the ASL-CBF, conventional + ASL-CBF and conventional + DSC-rCBF models achieved satisfactory performance, with AUC of 0.955, 0.968 and 0.960. While the DSC-rCBF model alone showed an AUC of 0.887 and decreased specificity (58.3%), and the conventional model demonstrated decreased AUC of 0.720 and specificity (33.3%). In this exploratory study, radiomics features derived from both ASL-CBF and DSC-rCBF showed promise in distinguishing RIBI from TR. Integrating perfusion features into conventional MRI enhanced model accuracy. These preliminary findings suggest ASL-based radiomics as a viable, non-contrast alternative; however, these results should be interpreted with caution due to the modest validation sample size and single-center design. They require validation in larger, prospective, multi-center cohorts before clinical implementation.

Tags

Topics