Quantum AI for psychiatric diagnosis: enhancing dementia classification with quantum machine learning

Le résumé fourni par la source

Early detection of dementia is a key requirement for effective patient management. Therefore, classification of dementia is pertinent and requires a highly accurate methodology. Deep learning (DL) models process immense amounts of input data, whereas quantum machine learning (QML) models use qubits and quantum operations to enhance computational speed and data storage through algorithms. QML is a research domain that investigates the interactions between quantum computing concepts and machine learning. A quantum computer reduces training time and uses qubits that play a vital role in learning complex imaging patterns, unlike convolutional kernels. The proposed study focused on imaging data and QML because they are more efficient and accurate than ML/DL for practical applications. Therefore, a hybrid quantum-classical convolutional neural network (QCNN) is proposed that integrates both quantum and classical learning paradigms. In the proposed framework, MRI images are pre-processed through resizing and normalization, followed by the extraction of a region of interest (ROI) from the center of each image. Within the ROI, a 2×2 patch is passed to a quantum circuit, where pixel values are encoded as qubits using rotation gates (RY). A parameterized quantum circuit (PQC) with entangling layers computes expectation values to generate a quantum feature map, which is then utilized as input to the classical CNN. To further improve generalization, a knowledge distillation (KD) framework is employed, where a teacher model (a deeper CNN with high representational capacity) guides a student model (the QCNN), transferring soft-label information via a temperature-scaled softmax. This setup enables the student model to learn more discriminative features while maintaining efficiency. Comprehensive experiments are conducted on benchmark ADNI-1, ADNI-2, and OASIS-2 MRI datasets, and results are reported both with and without KD. Without KD, the QCNN achieves strong performance with accuracies of 0.9523 (ADNI-1), 0.9611 (ADNI-2), and 0.9412 (OASIS-2). With KD, the student model demonstrates enhanced sensitivity to challenging classes, achieving an accuracy of up to 0.9978, surpassing state-of-the-art approaches. Combining quantum feature extraction with teacher-student knowledge transfer yields a scalable and highly accurate framework for dementia classification in clinical practice.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.