# mpMRI‐based MGMT methylation status prediction for glioblastoma through off‐the‐shelf deep features: A multi‐dataset feasibility study

## TL;DR

This study explores using MRI scans and deep learning to predict MGMT methylation status in glioblastoma, offering a non-invasive alternative to traditional methods.

## Contribution

A novel Siamese neural network approach using off-the-shelf deep features and delta T1 features for MGMT methylation prediction in glioblastoma.

## Key findings

- The method achieved an average AUC of 0.666 in predicting MGMT methylation status.
- Incorporating delta T1 deep features improved model performance confirmed through ablation studies.
- External validation showed reliable results with an AUC of 0.624 across different datasets.

## Abstract

O‐6‐methylguanine DNA methyltransferase (MGMT) promoter methylation status is a critical prognostic factor in glioblastoma. The aim of this study is to evaluate the feasibility of diagnosing MGMT status in a rapid, non‐invasive manner using multiparametric magnetic resonance imaging (mpMRI). The proposed method seeks to reduce reliance on stakeholders, thereby facilitating potential clinical applications in the future.

This study employed a Siamese neural network (SNN) as the backbone of the model to effectively leverage information from various mpMRI modalities. Off‐the‐shelf deep learning features extracted from pre‐trained networks was used to represent the information from mpMRI and adopted as the inputs of SNN. Delta deep features from T1 modality were integrated as additional branch of SNN to enhance model's performance. Finally, external validation was performed to increase the robustness of study. The proposed method was applied to one of the largest publicly available mpMRI datasets, comprising 585 participants, with an additional 81 samples used for external validation.

The proposed method achieved an average area under the curve (AUC) of 0.666 with a standard error of the mean (SEM) of 0.031, average precision of 0.591 (SEM 0.021), and average recall of 0.630 (SEM 0.064). In external validation, the method yielded an average AUC of 0.624 (SEM 0.022), precision of 0.674 (SEM 0.050), and recall of 0.810 (SEM 0.101).

The results demonstrate that our method outperforms existing approaches on a single‐CPU platform. Ablation studies confirmed the effectiveness of incorporating delta T1 deep features, while external validation confirmed the method's reliability across different datasets.