A 3D feature fusion model integrating multi-scale MRI feature for interpretable Glioblastoma prediction.

The existing predictive models often focuses on only one classification criterion, such as histological grade or isocitrate dehydrogenase (IDH) status. However, recent revisions to the WHO classification for glioblastoma multiforme (GBM), incorporating IDH mutation status alongside histology grade 4 classification, demands new predictive models capable of addressing both the characters defining new GBM status. The model functions as a black-box with limited interpretability, increasing the complexity of the predicting traits from biomedical images. A more holistic approach containing multi-scale imaging properties is required to characterize the highly heterogeneous nature of GBM. Here, we proposed a 3D feature fusion model (3D-FFM) to predict GBM status from 3D MRI data. It first extracts features from MRI images using various algorithms, including computer vision and image analysis techniques. The extracted features represent the physical characteristics of images and not abstract representations, helping to understand factors affecting GBM status. The developed framework integrates machine learning, recursive feature selection, and convolutional neural networks to build the final predictive model. Our findings unveil key radiomic features responsible for exhibiting significantly high predictive accuracy compared to existing methodologies, showcasing the potential of our hybrid machine learning approach in GBM diagnosis.