Unified CAD Framework For Brain Tumor Prediction And Segmentation Using Hybrid CNN-LSTM And 3D U-Net Architectures

Abstract

Brain tumors are among the most aggressive and life-threatening malignancies, requiring accurate detection and segmentation for effective treatment planning. Deep learning has significantly advanced medical imaging, but most studies focus on either tumor classification or segmentation, limiting their clinical utility. This paper extends our previous research on CNN-LSTM-based cancer prediction and U-Net-based glioma segmentation into a unified computer-aided diagnosis (CAD) framework. The proposed system first employs a 3D U-Net to localize tumor subregions from multi-modal MRI scans (BraTS 2020 dataset), generating voxel-wise masks for necrotic/non-enhancing core, peritumoral edema, and enhancing tumor regions. These segmented regions are then processed through a feature fusion strategy, where automatically extracted CNN features are combined with handcrafted features and classified using an LSTM-based network to predict tumor grade. Experimental evaluation demonstrates superior performance with a Dice score of 78%, mean IoU of 90%, and classification accuracy of 99.2%, surpassing existing standalone models. The integration of segmentation and classification reduces false positives, enhances interpretability, and provides a clinically relevant diagnostic tool. This study highlights the potential of end-to-end hybrid CAD systems in neuro-oncology and outlines pathways for incorporating explainable AI in future work.