Abstract

Deep learning-based methods have advanced medical image segmentation performance in recent years. However, these methods require further improvement and large amounts of annotated images to train. Image segmentation is mainly based on pixel-wise annotations, but obtaining annotated data is costly and stressful. Generative Adversarial Networks (GANs) are used to generate additional training data synthetically to avoid these costs and benefit from small datasets. Current mainstream GAN networks achieve high accuracy when trained on large, annotated datasets. Still, the potential annotated small datasets in medical image analysis have not been addressed much. This paper presents a novel approach to image segmentation that leverages GANs' powerful generative capabilities to address inherent challenges in segmentation tasks. Our method integrates a GAN framework in which the generator produces synthetic images, while CNN methods segment lung tumour images from both real and synthetic datasets. Afterwards, the results are compared to demonstrate the effect of data augmentation on lung tumour image segmentation. The experimental results show that our GAN-based Trans-UNet segmentation approach outperforms the UNet segmentation method, with the Trans-UNet on the synthetic data achieving increases of 0.6%, 0.7%, and 0.4% in Dice, PPV, ACC, and Sens, respectively, providing significant validation of the effectiveness of our approach. These results demonstrate the potential of GANs for addressing challenges in medical image segmentation and highlight the proposed method's improved accuracy and robustness.