Detection and Segmentation of Brain Tumor from MRI Image Signal using a Two-stage Deep Learning Framework

Abstract

Accurate detection and segmentation of brain tumors from magnetic resonance imaging (MRI) are essential for early diagnosis, treatment planning, and outcome monitoring in neuro-oncology. Manual assessment is labor-intensive, subjective, and prone to inter- and intra-observer variability, motivating the development of automated, AI-based solutions. This work proposes a two-stage deep learning framework for automated analysis of low-grade glioma (LGG) from FLAIR MRI images, consisting of a convolutional neural network (CNN) classifier for tumor detection followed by a Densely Convolutional Spatial Attention Network (DCSAN) for precise tumor segmentation. The framework is evaluated on the publicly available LGG Segmentation Dataset from The Cancer Imaging Archive (TCIA), comprising approximately 3,900 FLAIR MRI slices from 110 patients with corresponding expert-annotated masks. Our proposed CNN-based detection model achieves roughly 94% classification accuracy for distinguishing normal from abnormal scans, while the DCSAN attains about 87% Intersection over Union (IoU) and 93% Dice Similarity Coefficient (DSC) for tumor segmentation on a held-out test set, outperforming conventional U-Net–style and a Squeeze-and-Excitation based channel attention network (SEU-Net) architecture. A Tkinter-based Graphical User Interface (GUI) was also developed to demonstrate real-time image loading, prediction, and visualization of segmentation results. The proposed system delivers fast, reproducible, and high-fidelity tumor delineation, providing a promising decision-support tool to assist radiologists and neurosurgeons in clinical workflows.