TISSUE-SPECIFIC VIRTUAL STAINING USING TRANSFER LEARNING BASED ON GENERATIVE ADVERSARIAL NETWORK

Keywords: Transfer Learning, Virtual Staining, Generative Adversarial Network, Deep Learning, Computational Pathology

TL;DR: This paper demonstrates the advantages of transfer learning with a GAN model for virtual staining of histopathological whole-slide images.

Abstract: Generative adversarial networks (GANs) enable virtual staining of histopathological images, but training from scratch is costly and data-intensive. To address this, transfer learning is applied using a DensePix2Pix with pre-trained weights on two tissue types (kidney and spleen). Evaluation with SSIM, PSNR, PCCR, and MSE shows improved image quality, reduced training time, and greater resource efficiency compared to baseline models. While transfer learning proves effective with limited datasets, challenges in domain adaptation and generalization across tissues remain, underscoring the need for fine-tuning and hybrid approaches in future medical imaging applications.

Serve As Reviewer: ~Pekka_Ruusuvuori1

Submission Number: 28