AI-Based Brain MRI Segmentation for Early Diagnosis and Treatment Planning of Low-Grade Gliomas in the USA

Abstract

The detection of brain tumors in the USA is a complex task that requires high accuracy from imaging modalities. While it's true that many early-stage brain tumors can be managed effectively, they are often more aggressive and more challenging to treat than their higher-grade counterparts, ultimately leading to a fatal outcome with an average survival time of just 7 years after diagnosis. Therefore, these types of tumors must be accurately identified from MRI images, which are the most effective tool for diagnosing brain abnormalities. We have developed two deep-learning convolutional neural network models, U-Net and DeepLab, to segment brain MRI scans. We apply image segmentation techniques, which cluster the parts of the brain images into tumor or nontumor areas. To assess the effectiveness of our segmentation algorithm, we employ a widely recognized and reliable measure known as the Dice coefficient. The Dice coefficient objectively assesses the similarity between the predicted segmentation results and ground truth data. Using the Dice coefficient, we can better understand how well our algorithm captures the complex nuances of the image data. Our dataset is a valuable resource for brain MRI segmentation tasks. It comprises images from The Cancer Imaging Archive (TCIA), which provides high-quality imaging of 110 patients with lower-grade gliomas included in the broader data collection from The Cancer Genome Atlas (TCGA). These patient-specific brain MRI scans are accompanied by manually created fluid-attenuated inversion recovery (FLAIR) masks, allowing for detailed segmentation and abnormality detection.