# Deep learning-based prediction of TERT mutation status from MRI for glioma molecular subtyping

## TL;DR

This study uses MRI scans and deep learning to predict TERT mutations in glioma patients, helping classify tumors non-invasively for better treatment planning.

## Contribution

A deep learning model (RegNet) is developed and validated for non-invasive prediction of TERT mutation status in gliomas using MRI.

## Key findings

- RegNet achieved an accuracy of 0.7742 and AUC of 0.7182 in predicting TERT mutation status from MRI.
- RegNet outperformed other models like MobileNet and ShuffleNet in capturing TERT mutation-related imaging features.
- The model offers a non-invasive alternative to tissue-based diagnostics for glioma molecular subtyping.

## Abstract

This study aimed to develop and validate a deep learning model based on preoperative MRI to non-invasively predict Telomerase Reverse Transcriptase (TERT) promoter mutation status in glioma patients.

A retrospective cohort of 100 patients with histologically confirmed high-grade glioma was included. Regions of interest (VOIs) were manually annotated on contrast-enhanced T1-weighted MRI sequences by senior radiologists. Five deep learning models (RegNet, GhostNet, MobileNet, ResNeXt50, ShuffleNet) were trained and evaluated using accuracy, precision, recall, and F1-score. The dataset was split into training (80%) and internal validation (20%) sets.

RegNet achieved the highest performance with an accuracy of 0.7742, recall of 0.8704, precision of 0.7163, and F1-score of 0.7023. It demonstrated superior ability to capture imaging features associated with TERT mutations compared to other models. The area under the ROC curve (AUC) for RegNet was 0.7182, indicating moderate discriminative power.

The RegNet model effectively predicts TERT promoter mutation status from routine MRI, offering a non-invasive tool for preoperative molecular subtyping of glioma. This approach may facilitate personalized treatment planning and address limitations of invasive tissue-based diagnostics. Further validation with multi-center data is warranted to enhance clinical applicability.