IRU-Net: Brain Tumor Detection in MRI Images Using the End-to-End Inception Residual UNet Model

Nowadays, brain tumors are prevalent conditions affecting millions of people worldwide. Detecting the brain tumor accurately in its early phases by brain MRI images is important for diagnosing the disease and avoiding mortality. The rise of deep learning techniques over the past few years has led to effective tools for the diagnosis of medical illnesses, predominantly for the segmentation of brain tumors. Unet is considered one of the most effective end-to-end convolutional neural networks for biomedical image segmentation. This article introduces an Inception pre-trained network on ImageNet with residual connections based on the UNet architecture (called IRU-Net) for brain tumor segmentation. IRU-Net enhances the ability to integrate contextual information by using the GoogLeNet network and adding a residual connection as an encoder part. Additionally, the Inception module that incorporates residual connections is used as a decoder part, enabling the model to capture multi-scale features and ensure efficient gradient flow. The experiments conducted on two different size datasets obtained from Kaggle, namely Brain Tumor Segmentation (BraTS2020) and LGG MRI segmentation (LGG) which is smaller than BraTS2020. The presented results show the influence of pre-training on the network's performance using datasets of varying sizes. The experimental results demonstrate that the proposed IRU-Net architecture outperforms both the traditional UNet and other models that are based on pre-trained UNet..