Multi-Task Deep Learning Framework for Segmentation and Multi-Class Classification of Brain Tumors from MRI Images

Abstract

Brain tumors are among one of the most critical and life threatening diseases that can lead to severe consequences or even death if it is not detected and treated at an early stage. Accurate , timely and faster diagnosis of brain tumor is crucial for improving the surviving rates of patients. However, throughout the years manual methods of brain tumor analysis has been prevalent where radiologists goes through a painstaking process of scanning through numerous MRI (Magnetic Resonance Imaging ) images to identify the correct type of tumor/ This manual process is time consuming and often prone to human error. In order to address these challenges, deep learning based specially convolutional neural network based methods have gained increasing attention. Most conventional methods developed using deep learning architecture and convolutional neural network architecture mostly focuses one of the two aspects : either classification or semantic segmentation. In our work, we propose a multi-task deep learning framework that integrates both segmentation and multi class classification for comprehensive brain tumor analysis. The proposed framework employs Attention Residual U-Net model for precise tumor segmentation using the Figshare Brain Tumor Dataset, achieving a dice coefficient of 0.80 indicating strong overlap between predicted mask and ground truth tumor masks. The segmented tumor masks are subsequently concatenated with the corresponding MRI images to generate a two channel image highlighting the tumor region and providing classifier with spatial and contextual feature information. These images are then used to train an InceptionV3 classifier based multi class classification model to differentiate among three tumor categories meningioma, glioma and pituitary. The classification framework incorporated with the two channel image achieved an accuracy of 93.23%, outperforming the conventional classification models which were trained using only the raw MRI images. The integration of segmentation guided enhancement within the classification pipeline significantly improved the model’s ability to focus on tumor regions, leading to better performance. Our proposed model serves as an end to end solution for automated brain tumor analysis, capable of performing both semantic segmentation to recognize tumor region as well as classify and identify the correct tumor type. This process provides a reliable and scalable approach suitable for real world implementation and clinical applications.