Hybrid Quantum and CNN Framework for 3D Brain Tumor Segmentation

Abstract

The purpose of diagnosis, treatment planning, disease monitoring and precise evaluation of brain tumours using 3D multi-modality magnetic resonance imaging (MRI) is essential. Existing 3D convolutional neural networks struggle to segment tumor sub-regions that differ in character without affecting their computation and robustness. A hybrid quantum classical CNN is proposed in this paper based on a lightweight variational quantum circuit and a residual 3D convolutional neural network for multi-class brain tumor segmentation. The evaluation of the framework was carried out using the BraTS 2020 dataset, which included T2, T1ce, T1, and FLAIR MRI techniques. Preprocessing techniques include Skull removal, N4 bias field correction, brightness normalisation, isotropic resampling, patch extraction, and data augmentation. Features acquired from the encoder in the 2D image are improved using angleencoding (i.e., the angles of the input pixels) in a quantum circuit consisting of parameterized rotations and controlled NOT entangling gates. The quantum-classical features are fused to enable voxel-wise classifications as background, necrotic/non-enhancing tumor core, peritumoral edema, and enhancing tumor. The segmentation consistency can be improved through morphological refinement and connected component analysis. The optimized model employs a mix of cross-entropy loss and dice and is evaluated with measures in cluding Dice similarity coefficient, F1-score, accuracy, recall, intersection over union, sensitivity, specificity, Hausdorff distance, and accuracy. Results yield 90.83% accuracy with a 89.52% Dice score. show practical hybrid quantum-classical segmentation performance.