DEVELOPMENT OF HYBRID U-NET MODEL COMBINING CLASSICAL CONVOLUTIONAL LAYERS WITH QUANTUM CIRCUITS FOR SEGMENTATION OF LOWER-GRADE GLIOMA IN MAGNETIC RESONANCE IMAGES

ABSTRACT Lower-grade gliomas (LGG) are slow-growing brain tumors requiring accurate segmentation in magnetic resonance imaging (MRI) for effective diagnosis and treatment planning. Classical U-Net models face limitations including imprecise boundary definition, high computational requirements, and overfitting on limited datasets. This study develops and deploys a hybrid U-Net model integrating classical convolutional layers with variational quantum circuits (VQCs) to enhance segmentation accuracy and computational efficiency. The quantum component leverages superposition and entanglement principles to capture complex patterns with fewer parameters. Trained on the Kaggle_3m MRI dataset (3,929 paired images), with single MRI modalities, the hybrid model achieved consistent improvements over classical U-Net: Dice score increased from 0.835 to 0.850 (1.5% improvement), IoU from 0.718 to 0.739 (2.1% improvement), and accuracy from 0.912 to 0.925 (1.3% improvement). The quantum circuits required approximately 70% fewer trainable parameters while demonstrating superior tumor boundary delineation. A Streamlit web application was developed for real-time MRI upload and tumor visualization, enabling practical clinical deployment. These results validate the potential of hybrid quantum-classical architectures for medical image segmentation, offering enhanced performance with reduced parameter complexity and accessible deployment for healthcare practitioners.