Application and analysis of wavelet knowledge distillation for stain normalization in digital histopathology

Stain normalization has a critical importance in terms of minimizing colour differences and increasing the consistency of images, especially in histological images. The Pix2Pix model, which is one of the Conditional Generative Adversarial Network models, is based on the Wavelet Knowledge Distillation method, and it is aimed to compress complex models and make them efficient. The Wavelet Knowledge Distillation method carries out the knowledge transfer from the teacher model to the student model by decomposing it into high and low frequency components. High frequency components carry detailed structural information, while low frequency components contain general colour and shape information. The effect of teacher and student models on stain normalization was evaluated. As a result of the experiments carried out using the MITOS-ATYPIA dataset, which contains images of the same sample taken from two different scanners, it was observed that the student model of the Wavelet Knowledge Distillation method performs close to the teacher model and produces faster results. This method, which is especially promising for real-time applications, enables the effective application of stain normalization in Computer Aided Diagnosis systems. With the Wavelet Knowledge Distillation method, stain normalization is made more efficient and the performance of the model is improved by effectively separating high and low frequency components according to the model size.