Abstract—Thyroid nodules are a type of lesion, which doctors often need advanced diagnostic tools to detect and conduct follow-up diagnoses. Supervised deep learning techniques, particularly Generative Adversarial Networks (GANs), have been used to extract essential features, detect nodules and generate thyroid masks. However, these approaches suffer significant challenges in obtaining training data due to the high cost of identifying the cancer area and mode collapse during training. Therefore, this study proposed an improvement to one GAN model, namely, the pixel-to-pixel (pix2pix) model, for thyroid nodule segmentation, where the generator was incorporated with a supervised loss function to address instabilities during GAN training. The model used a generator with an encode-decoder structure inspired by U-Net architecture to produce the mask. The discriminator of the model consists of a multilayered Convolutional Neural Network (CNN) to compare the real and generated masks. In addition, three loss functions, namely, binary cross-entropy loss, soft dice loss and Jaccard loss, combined with loss GAN were used to stabilise the GAN model. Based on the results, the proposed model achieved 97% detection accuracy of the cancer area from the ultrasound thyroid nodule images and segmented it using the stabilised model with a generator loss function value of 0.5. In short, this study showed that the improved pix2pix model produced greater flexibility in nodule segmentation accuracy compared with semisupervised segmentation models.
 
Keywords—thyroid nodules segmentation, ultrasound image, deep learning, generative adversarial networks, pix2pix, loss function


Copyright © 2025 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).