Enhanced Deep Brain Tumor Segmentation with Medical Image Signal of MRI

The segmentation of brain tumors presents significant challenges due to their diverse morphologies, infiltrative nature, and unclear boundaries. MRI images, as one of the most effective medical image signals, often suffer from intensity inhomogeneities, which vary across different devices and times, contributing to variability among patients. The majority of the tissue in these images is non-tumorous, leading to class imbalance. To address these challenges, we first employed convolutional neural networks, specifically fine-tuning an inception-v3 model, to determine the presence of tumors. This approach outperformed other pre-trained models, achieving an accuracy of 97.95%, significantly surpassing alternatives such as Xception and VGG16 by margins ranging from 0.25% to 4.2%. For tumor segmentation, we proposed a Res-UNet model to overcome the limitations of U-Net, such as gradient vanishing and low feature utilization. The Res-UNet achieved a Dice coefficient of 0.9457 and an IoU of 0.8971, outperforming various U-Net variants with improvements in Dice coefficient by 0.0835 to 0.7160 and in IoU by 0.1381 to 0.7669. These results validate the effectiveness of our approach, leading to a more promising brain tumor detection and segmentation system.