Tracking Therapy Response in Glioblastoma Using 1D Convolutional Neural Networks
Sandra Ortega-Martorell1, Ivan Olier1, Orlando Hernandez2
1Data Science Research Centre, Liverpool John Moores University, Liverpool L3 3AF, UK.
Cancers
|August 12, 2023
Summary
Machine learning with magnetic resonance spectroscopic imaging (MRSI) accurately identifies glioblastoma treatment response. This approach offers improved early detection of tumor relapse, aiding in therapy assessment.
Area of Science:
- Neuro-oncology
- Medical imaging
- Machine learning
Background:
- Glioblastoma (GB) is a difficult-to-treat brain tumor with frequent relapse.
- Magnetic resonance imaging (MRI) with Response Assessment in Neuro-Oncology (RANO) criteria is used for therapy response evaluation.
- Pseudoprogression and pseudoresponse phenomena complicate early assessment, and magnetic resonance spectroscopy (MRS/MRSI) is underutilized.
Purpose of the Study:
- To explore the potential of MRSI combined with machine learning for improved glioblastoma therapy response assessment.
- To validate a novel methodology using preclinical models.
Main Methods:
- Utilized MRSI in combination with one-dimensional convolutional neural networks (1D-CNNs).
- Optimized and validated the method in preclinical glioblastoma (GL261-bearing mice) models.
- Employed Grad-CAM for model explainability and visualization.
Main Results:
- 1D-CNN models accurately identified normal brain, unresponsive tumor, and treatment-responsive tumor regions.
- Generated color-coded maps showing N, T, and R regions achieved high accuracy (Dice scores).
- The proposed method outperformed previous approaches in accuracy and explainability.
Conclusions:
- The developed methodology offers enhanced opportunities for glioblastoma therapy response assessment.
- This approach may provide earlier indications of tumor relapse stages.
