Genetic Programming for Feature Learning and Feature Construction in Glioma Survival Prediction
Kunjun Chen · 2024
MRI imaging plays a vital role in the initial tumor screening process and also aids in constructing a high-dimensional feature space for survival prediction tasks through the combination of sequences and tumor subregions. Traditional GP methods applied to feature construction treat features as part of the terminal set, which limits their search capabilities. Combining feature learning with GP allows for the direct use of image data, albeit dependent on predefined extraction functions. Our goal is to design a FLFC method that integrates feature learning and construction to automatically locate ROIs, extract features, construct new features, and match classifiers in 3D MRI images. To this end, we have introduced a new GP structure with an interpretable function set. To verify the algorithm, experiments were performed using the publicly available BraTS 2020 dataset. Ours achieved an average accuracy of 94.14% in the binary classification task (differentiating between HGG and LGG) and 72.49% in the four-class task (within HGG subtypes). These results underscore the efficacy of the FLFC method in harnessing the complexity of MRI data for meaningful survival prediction, illustrating its potential as a robust tool in medical imaging analysis and cancer prognosis.