Application of a machine learning method to whole brain white matter injury after radiotherapy for nasopharyngeal carcinoma

Abstract

BackgroundThe purpose/aim of this study was to 1) use magnetic resonance diffusion tensor imaging (DTI), fibre bundle/tract-based spatial statistics (TBSS) and machine learning methods to study changes in the white matter (WM) structure and whole brain WM network in different periods of the nasopharyngeal carcinoma (NPC) patients after radiotherapy (RT), 2) identify the most discriminating WM regions and WM connections as biomarkers of radiation brain injury (RBI), and 3) supplement the understanding of the pathogenesis of RBI, which is useful for early diagnosis in the clinic.MethodsA DTI scan was performed in 77 patients and 67 normal controls. A fractional anisotropy map was generated by DTIFit. TBSS was used to find the region where the FA differed between the case and control groups. Each resulting FA value image is registered with each other to create an average FA value skeleton. Each resultant FA skeleton image was connected to feature vectors, and features with significant differences were extracted and classified using a support vector machine (SVM). Next, brain segmentation was performed on each subject’s DTI image using automated anatomical labeling (AAL), and deterministic white matter fiber bundle tracking was performed to generate symmetrical brain matrix, select the upper triangular component as a classification feature. Two-sample t-test was used to extract the features with significant differences, then classified by SVM. Finally, we adopted a permutation test and ROC curves to evaluate the reliability of the classifier.ResultsFor FA, the accuracy of classification between the 0–6, 6–12 and > 12 months post-RT groups and the control group was 84.5, 83.9 and 74.5%, respectively. In the case groups, the FA with discriminative ability was reduced, mainly in the bilateral cerebellum and bilateral temporal lobe, with prolonged time, the damage was aggravated. For WM connections, the SVM classifier classification recognition rates of the 0–6, 6–12 and > 12 months post-RT groups reached 82.5, 78.4 and 76.3%, respectively. The WM connections with discriminative ability were reduced.ConclusionsRBI is a disease involving whole brain WM network anomalies. These brain discriminating WM regions and WM connection modes can supplement the understanding of RBI and be used as biomarkers for the early clinical diagnosis of RBI.

Machine-learning based MRI radiomics models for early detection of radiation-induced brain injury in nasopharyngeal carcinoma

“…d-f model 2 using features derived from medial temporal lobe, gray matter, and white matter respectively. g-i model 3 using features derived from medial temporal lobe, gray matter, and white matter respectively matter volume decreased markedly at acute and subacute stages after radiotherapy [21][22][23]. Lin J et al observed increased cortical thickness in NPC patients in the early period after radiotherapy [24].…”

Investigation of Radiation-Induced Toxicity in Head and Neck Cancer Patients through Radiomics and Machine Learning: A Systematic Review

“…We observed some interobserver discrepancies in the interpretation of specific RQS score items: the greatest were found for RQS score items “potential clinical applicability” and “open science and data,” even if other detected disagreements claim the need for more robust and easily interpretable methodological scoring systems for radiomic studies. Globally, higher standard deviations (4, 5) were observed in the studies by Leng, Abdollahi, and Pota [ 11 , 14 , 17 ].…”

“…In addition, the lack of follow-up scans to lend support to the authors' findings is a relevant limitation. The remaining 3 papers [ 11 , 13 , 15 ] were centered on magnetic resonance imaging (MRI) data. In a similar comparison with CT-based accuracy, the application of T1-weighted pretreatment MRI [ 13 ] was able to better detect the relationship between functional and nonfunctional parotid tissue and to improve the prediction ability of late xerostomia (AUC 0.83).…”

“…However, addressing MRI complex standardization is paramount. Taking all together, only one paper [ 13 ] reported on an external validation cohort and only 3 studies [ 11 , 12 , 17 ] can be considered in terms of “delta” radiomics. In this perspective, the analysis of intensity and texture features may contribute to the unresolved issue of replanning in HNSCC [ 20 ].…”

Interaction of Chemotherapy and Radiotherapy in Altering the Morphology of Subcortical Structures in Patients With Nasopharyngeal Carcinoma

“…For example, studies using structural MRI data have revealed that compared with patients before RT, patients with NPC have signi cantly decreased gray matter after RT, mainly involving the temporal lobe, hippocampus and the cerebellum, and that the gray matter decreases were dose-and time-dependent (Blanchard et al, 2015;Guo et al, 2018;Leng et al, 2017;Lin et al, 2017;Lv et al, 2014;Shi et al, 2018). Using diffusion tensor imaging (DTI), studies have found that RT could decrease fractional anisotropy and increase mean diffusivity of the white matter in cerebellum, temporal lobe, frontal lobe and parietal lobe in patients with NPC (Duan et al, 2016;Leng et al, 2017;Leng et al, 2019).…”