# A multitask framework based on CA-EfficientNetV2 for the prediction of glioma molecular biomarkers

## TL;DR

This paper introduces a multitask deep learning framework using MRI data to accurately predict glioma molecular biomarkers, improving non-invasive diagnosis and treatment planning.

## Contribution

A novel multitask framework combining CA-EfficientNetV2 with pseudolabel refinement and subnetworks for predicting glioma biomarkers.

## Key findings

- The TU-net subnetwork achieved 95.98% accuracy for IDH mutation prediction and 92.69% for MGMT promoter methylation.
- The framework outperformed other CNN-based approaches in predicting glioma molecular markers.
- The model uses a coordinate attention mechanism and optimized pseudolabels to enhance performance.

## Abstract

Glioma is the most common primary malignant tumor of the central nervous system. The mutation status of isocitrate dehydrogenase (IDH) and the methylation status of the O6-methylguanine-DNA methyltransferase (MGMT) promoter are key biomarkers for glioma diagnosis and prognosis. Accurate, non-invasive prediction of these biomarkers using MRI is of significant clinical value.

We proposed a novel multitask deep learning framework based on Coordinate Attention-EfficientNetV2 (CA-EfficientNetV2) to simultaneously predict IDH mutation and MGMT promoter methylation status based on MRI data. Initially, unlabeled MR images were annotated using K-means clustering to generate pseudolabels, which were subsequently refined using a Vision Transformer (ViT) network to improve labeling accuracy. Then, the Fruit Fly Optimization Algorithm (FOA) was employed to assign optimal weights to the pseudolabeled data. The CA-EfficientNetV2 model, integrated with a coordinate attention mechanism, was constructed. The multitask framework comprised three independent subnetworks: T2-net (based on T2-weighted imaging), T1C-net (based on contrast-enhanced T1-weighted imaging), and TU-net (based on the fusion of T2WI and T1CWI).

The proposed framework demonstrated high performance in predicting both IDH mutation and MGMT promoter methylation status. Among the three subnetworks, TU-net achieved the best results, with accuracies of 0.9598 for IDH and 0.9269 for MGMT, and AUCs of 0.9930 and 0.9584, respectively. Comparative analysis showed that our proposed model outperformed other convolutional neural network (CNN) - based approaches.

The CA-EfficientNetV2-based multitask framework offers a robust, non-invasive method for preoperative prediction of glioma molecular markers. This approach holds strong potential to support clinical decision-making and personalized treatment planning in glioma management.