Wavelet Knowledge Distillation for Efficient Stain Normalization in Digital Histopathology

In digital histopathology, color standardization, known as stain normalization, is widely used in computer-aided diagnosis (CAD) systems. This study details the adaptation and implementation of the Wavelet Knowledge Distillation (WKD) method to CAD systems. The proposed method focuses on knowledge transfer between the teacher and student models within a specially designed Pix2Pix Generative Adversarial Network (GAN) for stain normalization in histopathology images. The student model, guided by the knowledge transferred by the teacher model using wavelet-based feature extraction, significantly improves the accuracy of stain normalization, which is crucial for preserving histological details and image quality. The WKD method has demonstrated high performance on the publicly available paired MITOS-ATYPIA dataset, outperforming state-of-the-art methods. Using the same settings, the Teacher model achieved a PSNR of 25.559, SSIM of 0.934, and RMSE of 7.270. Additionally, the student model used in the method yielded better results in the Fréchet Inception Uzaklığı (FID) metric compared to the teacher and baseline models.