Brain Tumor Segmentation and Classification using a Hybrid CNN-LSTM Model

Abstract

Brain image classification and segmentation are challenging tasks in the field of medical image analysis. Most of the traditional approaches have largely treated these two tasks in isolation; however, in clinical practice, they are deeply connected to each other. The aim of this project is to propose and implement a hybrid model capable of performing brain tumor segmentation and grade classification simultaneously; the core of this proposed model integrates a custom-built Convolutional Neural Network (CNN) designed to extract rich spatial features from multi-modal MRI volumes with a Bidirectional Long Short-Term Memory (BiLSTM) network that captures the contextual relationships between consecutive brain slices. All experiments are conducted on the BraTS 2020 benchmark dataset, a widely adopted standard in the medical imaging community. To situate the proposed model's performance in a meaningful context, it is compared with two baselines: a ResNet34-LSTM and a DenseNet201-LSTM. The proposed model achieves a mean Dice score of 0.7326 across tumor sub-regions (Whole Tumor: 0.8862, Tumor Core: 0.8372, Enhancing Tumor: 0.7813) and an AUC of 0.9541 for glioma grading, with a Balanced Accuracy of 0.9130. This work demonstrates that spatial feature extraction and sequential context modeling are not competing paradigms but natural complements and that a carefully designed hybrid architecture can bring both tasks under one roof without sacrificing performance on either.