A Deep Learning Methodology for Differentiating Glioma Recurrence from Radiation Necrosis using Multimodal MRI

Algorithm Development and Validation

Yang Gao, Xiong Xiao, Bangcheng Han, Guilin Li, Xiaolin Ning, Defeng Wang, Weidong Cai, Ron Kikinis, Shlomo Berkovsky, Antonio Di Ieva, Liwei Zhang, Nan Ji, Sidong Liu*

*Corresponding author for this work

Research output: Contribution to journalArticle


Background: The radiological differential diagnosis between tumor recurrence and radiation-induced necrosis (i.e., pseudoprogression) is of paramount importance in the management of glioma patients. Objective: This research aims to develop a deep-learning methodology for automated differentiation of tumor recurrence from radiation necrosis based on routine MRI scans. Methods: In this retrospective study, 146 patients who underwent radiation therapy after glioma resection and presented with suspected recurrent lesions at the follow-up MRI examination were selected for analysis. Routine MRI scans were acquired from each patient including T1, T2, and Gadolinium-Contrast-Enhanced T1 sequences. Of those cases, 96 (66%) were confirmed as glioma recurrence on post-surgical pathological examination, whilst 50 (34%) were diagnosed as necrosis. Five state-of-the-art deep neural network (DNN) models were applied to learn radiological features of gliomas and necrosis from MRI scans and to classify the lesions at single-modal, multimodal and subject levels. Sensitivity, specificity, accuracy and AUC were used to evaluate performance of the models. Preoperative diagnostic performance of the models at subject level was also compared to that of 5 experienced neurosurgeons. Results: DNN models based on multimodal MRI outperformed single-modal models with an average sensitivity of 0.876+/-0.035, specificity of 0.733+/-0.062, accuracy of 0.827+/-0.03, and AUC of 0.863+/-0.032 on image-wise classification. When these DNN models were evaluated on a subject basis by aggregating the classification results of the subject’s image stack, the performance further improved to a sensitivity of 0.958+/-0.024, specificity of 0.8+/-0.071, accuracy of 0.904+/-0.029, and AUC of 0.934+/-0.022, which was significantly better than the tested neurosurgeons (P=0.018 in sensitivity, P<0.001 in specificity, P=0.003 in accuracy, respectively). Conclusions: DNN offer a useful computational tool for the differential diagnosis between recurrent gliomas and necrosis, achieving high performance on routine MRI scans. The proposed method does not depend on lesion segmentation or handcrafted features and therefore may achieve a high clinical applicability.
Original languageEnglish
JournalJMIR Medical Informatics
Publication statusSubmitted - 2 May 2020

Fingerprint Dive into the research topics of 'A Deep Learning Methodology for Differentiating Glioma Recurrence from Radiation Necrosis using Multimodal MRI: Algorithm Development and Validation'. Together they form a unique fingerprint.

  • Cite this