Multi-View Learning for Addressing Unbalanced/Partial Data Modalities in Glioblastoma Survival Prediction

Overview

Glioblastoma (GBM) is an aggressive form of brain tumor with a poor prognosis. Accurate survival prediction is crucial for personalized treatment planning. Our study leverages multi-modal learning techniques to handle missing or unbalanced data modalities in Glioblastoma survival prediction.

Motivation

Survival prediction in Glioblastoma patients is challenging due to:

• Heterogeneous data: Imaging, pathology, molecular, and clinical features vary across patients.
• Missing modalities: Not all patients have complete data, making traditional deep learning models unreliable.
• Interpretability issues: Black-box models lack transparency, reducing their adoption in clinical settings.

Approach

We propose a multi-view learning framework that integrates incomplete data modalities to enhance survival prediction. Our methodology includes:

• Data preprocessing and feature extraction across MRI, histopathology, and clinical data.
• Multi-modal deep learning models trained on incomplete data.
• Interpretability techniques for model transparency and clinical trust.

Data Sources

We utilized publicly available datasets from The Cancer Imaging Archive (TCIA):

• MRI & Radiomic Data: Download (69GB, 630 participants, 10645 files in NIfTI format)

  • Includes MRI scans, segmented tumor images, and radiomic features.
  • Additional details: DOI Link

• Histopathology Images: Download (34 patients, 71 images)

  • Multi-parametric MRI scans from UPENN-GBM collection.
  • More details: DOI Link

Methods

Data Preprocessing

• Converted DICOM to NIfTI for MRI images.
• Extracted tumor-specific features using radiomics principles.
• Tiled histopathology images into smaller patches to improve computational efficiency.
• Performed Exploratory Data Analysis (EDA) on clinical data, handling missing values and feature correlations.

Results & Findings

• The multi-view model performed better than single-modality models in survival prediction.
• The interpretability approach helped identify key biomarkers for prognosis.
• Future work includes expanding datasets and improving generalization.

Research References

Our work is supported by recent research papers:

1. Glioblastoma: An Update in Pathology, Molecular Mechanisms and Biomarkers
2. Epidemiology of Glioblastoma Multiforme
3. Survival Prediction using Machine Learning & Deep Learning
4. MRI-Based Survival Analysis