Transforming Frontal to Lateral : Chest X-ray View Conversion Using CycleGAN

Abstract

This thesis proposes a technique for transforming between frontal and lateral views of chest X-ray images using the Cycle Generative Adversarial Network (CycleGAN) and its variants, which are a type of Generative Adversarial Network (GAN). The lack of labeled paired data for frontal and lateral chest X-ray image translation presents a significant challenge in medical imaging applications. To overcome this limitation, this model is implemented to learn the correlation between frontal and lateral views of chest X-ray images.

The proposed approach involves training two generators and two discriminators, enabling the CycleGAN model to learn the correlation between the two views for a practical application. Extensive experimentation is conducted using a dataset including frontal and lateral view lung images, demonstrating the effectiveness of the CycleGAN framework in achieving accurate image translation.

The results indicate that the trained CycleGAN model partially successfully translates realistic looking between frontal and lateral views of chest X-ray images. Quantitative evaluation metrics such as Peak Signal-to-Noise Ratio and Structural Similarity Index confirm the fidelity of the translated images compared to ground truth frontal and lateral view images.

Overall, this research demonstrates the potential of CycleGAN in addressing the challenging task of frontal and lateral chest X-ray image translation, providing a valuable tool for medical imaging applications aimed at enhancing diagnostic accuracy and facilitating clinical decision-making