CLCT-GAN: Strong-Weak Contrastive Learning for Reconstructing CT Images from Radiographs

Generating CT images from radiographs has immense clinical potential, offering a novel approach to low-cost and low-radiation medical imaging. This method can significantly ease the workload of radiologists. Current machine and deep learning approaches mainly utilize fully supervised learning methods for CT image generation tasks. However, the feature representations learned in fully supervised settings are typically task-specific, dependent on large quantities of labeled data, and have limited generalizability. Addressing these challenges, we introduce a contrastive learning-based method for CT image generation, which includes two phases: pre-training and fine-tuning. The pre-training phase aims to learn feature representations from unlabeled radiographs. Specifically, we devised two simple yet efficient radiograph data augmentation methods, converting the original data into two related but different views. These are then input into an encoder module to learn discriminative feature representations. Labeled data is used to learn the medical image generation task during the fine-tuning phase. In evaluations across the LIDC-IDRI lung CT and IXI brain MRI datasets, our CLCT-GAN model exhibits not only outstanding performance in lung CT reconstructions but also showcases remarkable adaptability to brain MRI data from IXI, surpassing previous state-of-the-art models in these diverse medical imaging benchmarks.