Wavelet-Based Loss Function Approach for Synthetic IHC Generation from H&E in Breast Cancer Histopathology Images Meme Kanseri Histopatoloji Görüntülerinde H&E'den Sentetik IHC Üretimi Için Dalgacik Tabanli Kayip Fonksiyonu Yaklaşimi
34th Signal Processing and Communications Applications Conference, SIU 2026, İstanbul, Turkey, 7 - 10 July 2026, (Full Text)
- Publication Type: Conference Paper / Full Text
- Doi Number: 10.1109/siu71813.2026.11636639
- City: İstanbul
- Country: Turkey
- Keywords: Breast Cancer, Deep Learning, GAN, Histopathology, Virtual Staining, Wavelet Transform
- Yıldız Technical University Affiliated: Yes
Abstract
Immunohistochemistry (IHC) is vital for breast cancer diagnosis but costly and time-consuming compared to Hematoxylin and Eosin (H&E) staining. This study explores synthetic IHC generation (virtual staining) from H&E using deep learning. We analyze the impact of a Discrete Wavelet Transform (DWT)-based loss function on the texture and spectral consistency of synthesized images. Experiments on BCI, MIST, and HER2match datasets using Pyramid Pix2Pix, Pix2Pix, and CycleGAN show DWT loss significantly improves perceptual metrics like FID and KID. Specifically, applying DWT loss to the Pyramid Pix2Pix model on the BCI dataset increased PSNR by 3.04% (22.55 dB) and improved FID by 20.4% (118.83) compared to the baseline. These findings prove that frequency-based losses effectively mitigate the perception-distortion trade-off and preserve high-frequency texture details in virtual staining.