Räumlich basierte metabolische Klassifizierung durch Graph Attention Deep Learning

Metabolic analysis of tumor of the central nervous system is a key component of the preoperative diagnostic work-up in neuro-oncology. In recent years, there has been an increasing interest in deepening the prediction capabilities of preoperative radiologic tools to better understand tumor genetics. With the help of artificial intelligence, we aimed at connecting these two very important pillars.

Data were obtained by the Phase I SPORT Trial (DRKS00019855) including n=134 patients who underwent MR-Spectroscopy. We developed a graph attention network (GAT) model to predict MR spectroscopy outcomes by focusing on key connections betweenneighborhood metabolic patterns. The model uses explainable AI tools to highlight key relationships and provides interpretable results to understand which metabolic pattern drives the prediction of genetic tumor subgroups.

All 134 patients were classified accordingly to the 2021 WHO classification into 15 different tumor types. The GAT classified the tumor types in a 10-fold cross-validation with a mean accuracy of 98,7% (Precision 0.9892 Recall 0.9889 and F1-score 0.9890). Using computational AI tools such as attention and integrated gradients, we found that spatial patterns of metabolic changes were more important for accurate predictions than local spectra. Decreasing the size of neighborhood spectra significantly reduces the prediction accuracy.

Leveraging spatial relationships of metabolic alterations in CNS tumors significantly increases the accuracy of tumor classification in MR-Spectroscopy. Our novel model can provide explainable results in the diagnosis and interpretation of MR spectroscopy.