Genotype Characterization in Primary Brain Gliomas via Unsupervised Clustering of Dynamic PET Imaging of Short-Chain Fatty Acid Metabolism

The impact of genetics on the diagnosis, treatment, and survival outcomes for patients with brain glioma is significant. At present, isocitrate dehydrogenase (IDH) mutation, the key biomarker in brain glioma with considerably better-survival rates, lacks a distinct radiologic signature. In this study, we targeted the glioma specific mechanism involving short chain fatty acid (SCFA) transcellular flux (TF) for energy production using 18F-fluoropivalate (FPIA) PET tracer and used this information to characterize the genetic profile of 10 patients with brain gliomas (5 IDH-mutant and 5 wild-type). We discerned four unique SCFA metabolic profiles by applying k-means clustering to an average of $25202~(\pm ~14337$ ) time activity curves (TACs) extracted from dynamic 18F-FPIA PET scans. Using deep learning (DL), the TACs from the first two clusters accurately differentiated between mutant and wild-type gliomas ( $96.75\pm 3.24$ % accuracy, $0.96\pm 0.04$ AUC). The third cluster, the one with the lowest-FPIA SUV, showed the worst performance ( $23.67\pm 16.83$ % accuracy, $0.31\pm 0.17$ AUC), suggesting that only a subset of SCFA-TF profiles define the genetic status of the tumor. Finally, disregarding the heterogeneity of SCFA-TF significantly reduced our model’s effectiveness, with accuracies dropping to $67.40\pm 22.87$ % and $70.42\pm 16.25$ % when tested using static SUV PET data and the full range of FPIA TACs, respectively.