Detecting IDH and TERTp mutations in diffuse gliomas using H-MRS with attention deep-shallow networks.

Background: Preoperative and noninvasive detection of isocitrate dehydrogenase (IDH) and telomerase reverse transcriptase gene promoter (TERTp) mutations in glioma is critical for prognosis and treatment planning. This study aims to develop deep learning classifiers to identify IDH and TERTp mutations using proton magnetic resonance spectroscopy (H-MRS) and a one-dimensional convolutional neural network (1D-CNN) architecture.

Methods: This study included H-MRS data from 225 adult patients with hemispheric diffuse glioma (117 IDH mutants and 108 IDH wild-type; 99 TERTp mutants and 100 TERTp wild-type). The spectra were processed using the LCModel, and multiple deep learning models, including a baseline, a deep-shallow network, and an attention deep-shallow network (ADSN), were trained to classify mutational subgroups of gliomas. The Gradient-weighted Class Activation Mapping (Grad-CAM) technique was used to interpret the models' decision-making process.

Results: The ADSN model was the most effective for IDH mutation detection, achieving F1-scores of 93 % on the validation set and 88 % on the test set. For TERTp mutation detection, the ADSN model achieved F1-scores of 80 % in the validation set and 81 % in the test set, whereas TERTp-only gliomas were detected with F1-scores of 88 % in the validation set and 86 % in the test set using the same architecture.

Conclusion: Deep learning models accurately predicted the IDH and TERTp mutational subgroups of hemispheric diffuse gliomas by extracting relevant information from H-MRS spectra without the need for manual feature extraction.