5521. Unraveling the Challenges of MGMT Classification in Gliomas: A Study on IDH and MGMT Using MCP and MRI
Authors * Denotes Presenting Author
  1. Shahriar Faghani; Mayo Clinic - Rochester
  2. Mana Moassefi *; Mayo Clinic - Rochester
Recent research has demonstrated the potential of deep learning (DL) models in predicting genetic status based on medical imaging. Nevertheless, the extent to which DL can accurately predict the status of all genes remains uncertain. Notably, a growing body of literature suggests that the methylation status of the O-6-methylguanine-DNA methyltransferase (MGMT) gene promoter in glioma cannot be determined using MRI. To address this issue, the present study investigates the application of Mondrian Conformal Prediction (MCP) in the classification of isocitrate dehydrogenase (IDH) and MGMT in glioblastomas utilizing MRI. MCP assigns conformity scores to predictions, allowing the identification of certain and uncertain cases. By focusing on certain cases and disregarding the uncertain ones, it is anticipated that the model's performance is anticipated to improve if the previously observed low performance was due to uncertainty rather than the model's ability to distinguish between classes.

Materials and Methods:
The publicly available UCSF-PDGM dataset, which comprises MRI sequences and corresponding tumor genetic profiles, was utilized in this analysis. The dataset consisted of MGMT negative: positive samples in a ratio of 1:3.34 and IDH wildtype: mutant samples in a ratio of 1:3.81. The dataset was divided into five folds, stratified by labels at the patient level. Two 3D-Densenet121 models were trained on the T2 sequence for predicting IDH and MGMT. The efficacy of MCP in IDH and MGMT classification was evaluated.

The accuracy and AUC values for IDH classification were 0.83 and 0.88, respectively, and for MGMT prediction, they were 0.78 and 0.54. With the application of MCP, the accuracy for IDH prediction improved to 1, but no improvements were observed in MGMT classification.

The findings emphasize that MCP enhances model performance when the model can distinguish between different classes and further support the notion that current classifiers struggle to effectively capture MGMT signals from MRI data.