The following is a summary of “Breast Ultrasound Images Augmentation and Segmentation Using GAN with Identity Block and Modified U-Net 3+,” published in the October 2023 issue of Oncology by Alruily et al.
Breast cancer is one of the most common diseases affecting women today. Early detection can lead to more effective treatment and better outcomes. Researchers performed a retrospective study to develop a hybrid approach to augment and segment breast cancer images.
The study consisted of two main stages that were ultrasound image augmentation and segmentation. The augmentation stage used a generative adversarial network (GAN) with a nonlinear identity block, label smoothing, and a new loss function. The segmentation stage uses a modified U-Net 3+ model.
The study demonstrated that the hybrid approach outperformed other methods in both segmentation and augmentation steps for the same task. The modified GAN with nonlinear identity blocks showed superior performance compared to various ultrasound augmentation techniques, including speckle GAN, UltraGAN, and deep convolutional GAN. Similarly, the modified U-Net 3+ outperformed different U-Net architectures in the segmentation process. The GAN with nonlinear identity blocks achieved an inception score of 14.32 and a Fréchet inception distance of 41.86 during augmentation. In contrast, the GAN with identity had a smaller Fréchet inception distance (FID) and a higher inception score.
The study found that GAN-U-Net outperformed other GANs in ultrasound image augmentation and segmentation, achieving a Dice Score of 95.49% and an Accuracy of 95.67%.