About the Validity of Using DCGANs for Data Augmentation in Breast Thermography Segmentation

This study explores the use of Deep Generative Adversarial Networks (DCGANs) and U-Net Convolutional Neural Networks for generation and segmentation of infrared breast images. Due to the scarcity of available thermal data, we used DCGANs to generate new thermal images, thus increasing the amount of data for training. The network was trained with 10,000 seasons, resulting in realistic synthetic images. We developed two segmentation models based on U-Net: one using only real images and other using real images as well as the synthetic generated images to help in data augmentation. Comparing the results, we observed that the accuracy of the segmentation model increased from 35.2% to 90.4%, and the specificity increased from 0% to 97.6% after applying DCGAN. The validation scores indicated high accuracy and robustness of the developed model. This work contributes to the advancement of thermal image generation and segmentation techniques, offering a promising approach for the creation of augmented databases and applications in medical diagnostics.