Quantum neural networks for medical image analysis

Medical image analysis has benefited enormously from the rise of deep learning and increasing accuracy of neural network models to classify and segment images. The problem is however far from being solved where deep learning models are known to be very compute intensive and to require large amounts of annotated data, both of which can pose severe limitations to their adoption in production. To this end, several different approaches have been proposed to apply the power of quantum computing to neural networks and machine learning in general. While many of these approaches focus on fault tolerant machines, our team optimized quantum circuits to perform key linear algebraic operations on near term machines which are small and noisy. These operations, namely vector to vector and vector to matrix multiplication, form the foundation of common neural network architectures. Efficient encoding and quantum ansatz with guarantee on asymptotical scalability were found for orthogonal fully connected layers and quantum attention mechanism. Efficiency of the proposed methods was demonstrated on the MedMNIST, a collection of standardized medical image classification tasks with classical benchmark results, both using simulation and hardware experiments.