Abstract Intra-tumoral heterogeneity (ITH) complicates the diagnosis and treatment of glioma, partly due to the diverse metabolic profiles driven by genomic alterations. While multiparametric imaging captures spatial and functional variations enabling ITH characterization, it falls short in assessing the metabolic activities underpinning phenotypic differences. This gap stems from the challenge of integrating easily accessible, co-located pathology and genomic data with metabolic insights. This study presents a multi-faceted approach combining stereotactic biopsy with clinical open-craniotomy for sample collection, voxel-wise analysis of MR images, regression-based GAM, and whole-exome sequencing. The aim is to demonstrate the potential of machine learning algorithms to predict variations in cellular and molecular tumor characteristics. This retrospective study enrolled ten treatment-naïve patients with radiologically confirmed glioma who underwent multiparametric MR scans (T1W, T1W-CE, T2W, T2W-FLAIR, DWI) prior to surgery. During craniotomy, at least one stereotactic biopsy was collected from each patient, with the screenshots of sample locations saved for spatial registration to pre-surgical MR data. Whole-exome sequencing was performed on flash-frozen tumor samples, prioritizing five glioma-related genes: IDH1, TP53, EGFR, PIK3CA, and NF1. Regression was implemented with a GAM using a univariate shape function for each predictor. Standard ROC analyses were used to evaluate detection, with AUC calculated for each gene target and MR contrast combination. Mean AUC for five gene targets and 31 MR contrast combinations was 0.75±0.11, with individual AUCs as high as 0.96 for IDH1 and TP53 with T2W-FLAIR and ADC, and 0.99 for EGFR with T2W and ADC. An average AUC of 0.85 across the five mutations was achieved using the combination of T1W, T2W-FLAIR, and ADC. These results suggest the possibility of predicting exome-wide mutation events from non-invasive, in vivo imaging by combining stereotactic localization of glioma samples with a semi-parametric deep learning method. This approach holds potential for refining targeted therapy by better addressing the genomic heterogeneity of glioma tumors.
Read full abstract