Development of Prediction Models to Detect the Presence of MGMT Promoter Methylation for Prognosis of Brain Tumor

Abstract

Methylation of the MGMT promoter is a molecular marker of predictive relevance in brain tumors. Methylation of the MGMT promoter is related to an improved prognosis and may influence therapy decisions. According to the most recent Malaysian National Cancer Registry Report (MNCRR) for 2012–2016, there were 2097 cases of brain tumors in Malaysia overall, with 1117 cases involving men and 908 cases involving women. This research aims to create prediction models for detecting MGMT promoter methylation in brain tumors. So, a dataset from the BraTS (Brain Tumor Segmentation) challenge, which included MRI scans and clinical data for patients with brain tumors, was utilized and explored using Exploratory Data Analysis (EDA) and data visualization techniques. The dataset has 306 methylation cases identified as “1” and 276 unmethylated cases identified as “0”. The average number of scans of train data for modality per patient is between 127 and 171, which provides a wealth of information for pattern learning while the average number of scans of test data for modality per patient is between 124 and 165. In both sets, FLAIR has the least number of files while T2w has the highest number of files among them. Two models of deep learning approaches, ResNet50 and EfficientNetV2, were used to construct and form the prediction models. Various criteria were employed to evaluate the performance of the models. The T2w modality consistently achieved the highest validation accuracy, precision, recall, and F1-score for both the ResNet50 and EfficientNetV2 models. Specifically, the T2w modality achieved a validation accuracy of 0.9453 and a validation loss of 0.2417 for ResNet50. Furthermore, the brain tumor diagnosis interface utilized DICOM images from MRI scans to identify MGMTp methylation status, aiding in therapy effectiveness prediction. Pre-trained ResNet50 model on T2w images was used for classification. The interface displayed the original MRI image, predicted state, and treatment effectiveness indication, while promptly notifying users of invalid or inadequate data for accurate analysis.