Efficient segmentation and classification of the tumor using improved encoder-decoder architecture in brain MRI images

Abstract

Primary diagnosis of brain tumors is crucial to improve treatment outcomes for patient survival. T1-weighted contrast-enhanced images of Magnetic Resonance Imaging (MRI) provide the most anatomically relevant images. But even with many advancements, day by day in the medical field, assessing tumor shape, size, segmentation, and classification is very difficult as manual segmentation of MRI images with high precision and accuracy is indeed a time-consuming and very challenging task. So newer digital methods like deep learning algorithms are used for tumor diagnosis which may lead to far better results. Deep learning algorithms have significantly upgraded the research in the artificial intelligence field and help in better understanding medical images and their further analysis. The work carried out in this paper presents a fully automatic brain tumor segmentation and classification model with encoder-decoder architecture that is an improvisation of traditional UNet architecture achieved by embedding three variants of ResNet like ResNet 50, ResNet 101, and ResNext 50 with proper hyperparameter tuning. Various data augmentation techniques were used to improve the model performance. The overall performance of the model was tested on a publicly available MRI image dataset containing three common types of tumors. The proposed model performed better in comparison to several other deep learning architectures regarding quality parameters including Dice Similarity Coefficient (DSC) and Mean Intersection over Union (Mean IoU) thereby enhancing the tumor analysis.