2D Brain MRI Segmentation: U-Nets Versus Optimized DeepLab Models

Brain tumors can cause a variety of psychiatric symptoms. Early detection of these tumors can significantly improve the chances of treating or preventing other diseases like Alzheimer's, dementia, multiple sclerosis, and bipolar disorder. Our paper proposes a novel method using semantic segmentation to detect brain tumors in MRI scans. We compare two DeepLab architectures (DeepLab V3 and DeepLabv3 with ResNet-101 backbone) against a baseline U-Net model, all trained and validated on a dataset of 3,929 brain MRI images. DeepLabv3 with ResNet-101 achieved the highest accuracy (Dice coefficient of 0.8746), outperforming both U-Net (0.7517) and DeepLabV3 (0.8177). DeepLabv3 with ResNet-101 also achieved the highest training accuracy (Dice coefficient of 0.9605) and validation accuracy(0.8579). This research demonstrates that DeepLabv3 with ResNet-101 backbone surpasses the pre-existing methods for brain tumor detection in MRI scans.