Leveraging segmentation-guided spatial feature embedding for overall survival prediction in glioblastoma with multimodal magnetic resonance imaging

Junmo Kwon ¹, Jonghun Kim ², Hyunjin Park ¹

⁰Department of Electrical and Computer Engineering, Sungkyunkwan University, Suwon, 16419, South Korea; Center for Neuroscience Imaging Research, Institute for Basic Science, Suwon, 16419, South Korea.

Computer Methods and Programs in Biomedicine +

|

July 23, 2024

View abstract on PubMed

Summary

Accurately predicting glioblastoma survival is vital. Our novel segmentation-guided deep learning method improves overall survival prediction using multimodal MRI, outperforming baseline approaches.

Area Of Science

Neuro-oncology
Medical Imaging Analysis
Machine Learning

Background

Glioblastoma has a poor prognosis, with a five-year survival rate under 5%.
Accurate overall survival (OS) prediction is critical for effective glioblastoma treatment planning.

Purpose Of The Study

To develop a novel deep learning framework for predicting glioblastoma patient overall survival (OS) using multimodal magnetic resonance imaging (MRI).
To leverage brain tumor segmentation to enhance OS prediction accuracy.

Main Methods

A segmentation-guided regression method was proposed, utilizing a pre-trained brain tumor segmentation network.
The survival regression network was jointly trained, focusing on tumor voxels and suppressing background noise.
The framework employed a UNETR++ backbone and incorporated contrastive loss for improved performance.

Main Results

The proposed framework achieved a Dice score of 0.7910, Spearman correlation of 0.4112, and Harrell's concordance index of 0.6488.
The model demonstrated superior performance compared to baseline methods on BraTS and UCSF-PDGM datasets.
Ablation studies confirmed that pre-training the segmentation network and using contrastive loss significantly improved OS prediction metrics.

Conclusions

A joint learning framework utilizing a pre-trained segmentation backbone and brain tumor segmentation maps was successfully developed for OS prediction.
The model's spatial feature map enables a sliding-window approach, adaptable to varying input image resolutions and matrix sizes.
This approach offers a promising tool for improving survival prediction in glioblastoma patients.

Keywords:

Glioblastoma Joint learning Magnetic resonance imaging Overall survival prediction Semantic segmentation