Thyroid Cancer Ultrasound Image Segmentation Based on UNet Dimensional Variants and Transfer Learning GANs

UNet is a mainstream technique in medical image segmentation, and it is one of the most important semantic segmentation frameworks based on CNNs. With ongoing development, various UNet-related architectures and hybrid models have emerged, such as UNet++, Attention UNet, 3D UNet, and Res UNet. This paper focuses on the basic conceptual principles of UNet, the advantages and disadvantages of its variants on the performance of cancer image segmentation. Furthermore, the study explores the impact of transfer learning-based GAN on the performance of thyroid cancer image segmentation. Specifically, we examine the effects of different UNet dimensional variants (e.g., 2D, 2.5D, 3D) on segmentation performance, as well as the enhancement introduced by GAN-generated synthetic data through transfer learning. Experimental results show that the 2.5D UNet achieved the best performance on a dataset of 800 images, with a Test Dice score of 0.628 and a Jaccard index of 0.593. When the training set consisted of 600 original images and 400 GAN-generated synthetic images, the model trained with these augmented data achieved the best overall performance, with a Test Dice of 0.628 and a Jaccard of 0.525.