Adaptive Fine-tuning based Transfer Learning for the Identification of MGMT Promoter Methylation Status

Glioblastoma Multiforme (GBM) is an aggressive form of malignant brain tumorwith a generally poor prognosis. Treatment usually includes a mix of surgicalresection, radiation therapy, and akylating chemotherapy but, even with theseintensive treatments, the 2-year survival rate is still very low.O6-methylguanine-DNA methyltransferase (MGMT) promoter methylation has beenshown to be a predictive bio-marker for resistance to chemotherapy, but it isinvasive and time-consuming to determine the methylation status. Due to this,there has been effort to predict the MGMT methylation status through analyzingMRI scans using machine learning, which only requires pre-operative scans thatare already part of standard-of-care for GBM patients. We developed a 3DSpotTune network with adaptive fine-tuning capability to improve theperformance of conventional transfer learning in the identification of MGMTpromoter methylation status. Using the pretrained weights of MedicalNet coupledwith the SpotTune network, we compared its performance with two equivalentnetworks: one that is initialized with MedicalNet weights, but with no adaptivefine-tuning and one initialized with random weights. These three networks aretrained and evaluated using the UPENN-GBM dataset, a public GBM datasetprovided by the University of Pennsylvania. The SpotTune network enablestransfer learning to be adaptive to individual patients, resulting in improvedperformance in predicting MGMT promoter methylation status in GBM using MRIs ascompared to using a network with randomly initialized weights.