Lightweight Quantum-Convolutional Neural Network for Brain Tumor Classification

Abstract

Brain tumor classification is a critical task in medical imaging because early and accurate detection directly influences treatment decisions. This thesis develops and evaluates a lightweight hybrid quantum-classical convolutional neural network for brain tumor MRI classification. The model combines a shallow grayscale CNN encoder, a six-dimensional bottleneck, batch normalization, and a six-qubit parameterized quantum circuit with ring entanglement before final classification. It is evaluated in two settings: binary tumor versus non-tumor detection and fourclass classification of glioma, meningioma, pituitary, and non-tumor MRI images. To make the evaluation clearer, the hybrid model is compared with a matched lightweight classical baseline that uses the same preprocessing, encoder, bottleneck size, optimizer, data split, and training protocol, replacing only the quantum layer with a direct classical classifier. Experimental results show that the hybrid model improves binary accuracy from 85.08% to 93.29% and weighted F1-score from 0.8505 to 0.9327. In the four-class task, accuracy improves from 77.19% to 83.50% and weighted F1-score from 0.7612 to 0.8305. The results suggest that a compact parameterized quantum circuit can improve a small CNN bottleneck model under this controlled setup, while the remaining errors show that fine-grained tumor subtype classification is still more difficult than binary screening.