DeepSIT: Deeply Supervised Framework for Image Translation on Breast Cancer Analysis

Image translation networks are deep learning models that can convert an image from one domain to another while preserving the semantic content. These networks are helpful in the medical field for noise reduction, reconstruction, and modality conversion. In this work, we propose DeepSIT, a deeply supervised framework for image translation. DeepSIT is a conditional generative adversarial network composed of a deeply supervised U-Net generator network and four PatchGAN discriminator networks. The generator performs the translation task while the discriminators judge the quality of the generated images. Unlike other works, the generator has four output layers located in the final and intermediate layers of the network. Each output layer generates a synthetic image, which is evaluated using a pixel-wise L1 loss function. Furthermore, the four discriminator networks receive a predicted image from an output layer to judge the quality of the translation at different scales. A promising application of image translation is the generation of immunohistochemical (IHC) images from Hematoxylin and Eosin (HE) images for breast cancer diagnosis. The proposed framework is evaluated in the latter tasks using the BCI Image Generation Grand Challenge dataset. DeepSIT achieves first place in the post-challenge leaderboard with an average of 0.545 SSIM and 18.037 PSNR in the test set.