BioGraph-Net: Beyond Voxel-wise Analysis -A Self-Supervised Radiogenomic Platform via Dynamic Hyper-Graph Transformers for Non-Invasive Virtual Biopsy

Virtual biopsy (VB) is a novel approach that allows us to gather important information about tissue pathology and molecular details without the need for invasive procedures like surgery or needle sampling. It is worth noting that, it relies on analyzing medical imaging data with cutting-edge techniques like radiomics and artificial intelligence (AI). This marks a significant shift from the traditional method of physically retrieving tissue samples to using quantitative imaging biomarkers that can stand in for histopathology. When it comes to gliomas, understanding the molecular makeup; especially identifying Isocitrate Dehydrogenase (IDH) mutations; is crucial for tailoring effective cancer treatments. Unfortunately, surgical biopsies come with their own set of risks and can be limited by how well they sample the tissue. That’s where BioGraph-Net comes in. This innovative radiogenomic framework combines Swin-Transformers with Dynamic Hyper-Graph Neural Networks (DHGNN) to create a powerful tool for analysis. Using a Self-Supervised Masked Autoencoder (MAE) strategy, the model first learns the important anatomical features before it decodes complex genomic signatures from various types of MRI scans (T۱, T۱ce, T۲, FLAIR). After being validated on the BraTS ۲۰۲۵ benchmark, BioGraph-Net has shown impressive results, achieving a Dice Similarity Coefficient (DSC) of ۰.۹۴ and an Area under the curve (AUC) of ۰.۹۱ for predicting IDH status. This effectively paves the way for a clinically viable “Virtual Biopsy” approach, making strides toward more precise and less invasive cancer diagnostics.

کلیدواژه ها:

Virtual Biopsy ، Non-Invasive Molecular Profiling ، IDH Mutation ، Radiogenomics ، Swin-Transformer