High-Resolution Medical Image Generation with Leak-Prevention Mechanism using Quantum Transformer Learning

A high-performance computer-aided diagnosis (CAD) system enhances diagnostic accuracy, enabling early detection and treatment. However, limited medical image datasets and low-quality augmentation methods hinder training, affecting CAD performance. Data augmentation increases training data but introduces augmentation leaks, where distortions in augmented images appear in generated outputs. Maintaining invertibility in augmentation can prevent leaks, but ensuring this across various generative adversarial networks (GANs) and datasets remains challenging. Additionally, existing high-resolution image generation models rely on complex deep neural networks, making optimization difficult and leading to issues like blurred lesion boundaries and poor detection of micro-leakage and micro-vessels. To address these challenges, we propose quantum transformer learning (QTL) for high-resolution medical image generation with prevention to mitigate augmentation leaks. The QTL generator enhances detail generation in high-resolution images while the prairie dog optimization (PDO) algorithm is combined with the activation function to accelerate the model convergence. Furthermore, an auxiliary multi-head neural network (AMH-NN) classifies the augmentation process during GAN training, guiding the model to avoid generating images with unintended distortions. The effectiveness of the proposed approach is evaluated using the knee osteoarthritis and medical MNIST datasets. Extensive experiments demonstrate that the proposed model significantly reduces augmentation leaks and achieves superior image quality, attaining the lowest FID scores compared to state-of-the-art GAN models. The proposed QTL+AMH-NN model achieves 22.4% improvement in FID values for the Knee Osteoarthritis dataset and 18.6% enhancement for the Medical MNIST dataset compared to previous approaches. The most notable gain is observed in Breast-MRI (86.3%), demonstrating the model’s capability to generate high-quality medical images with minimal augmentation artifacts.