A Hybrid Generative Adversarial Network for High-Fidelity Image Super-Resolution in Digital Pathology

International Journal of Computer Science and Technology | Vol. 3, No. 6, June 2022 | pp. 41–48

Research Article

Title: A Hybrid Generative Adversarial Network for High-Fidelity Image Super-Resolution in Digital Pathology

Names of Authors: Jean Dupont¹, Marie Laurent², and Pierre Mercier³

Authors’ Affiliations:
¹Laboratory of Computer Science and Systems, Aix-Marseille University, Marseille, France
²Department of Information Technology, Télécom Paris, Palaiseau, France
³Institute of Digital Health, University of Montpellier, Montpellier, France

Abstract: High-resolution digital pathology scans are crucial for accurate tumor grading, tissue structure evaluation, and cellular anomaly identification. However, clinical scanning hardware is limited by optical constraints and high processing costs, often yielding histopathological images with lower pixel resolution that can obscure critical diagnostic markers. This paper presents a hybrid Generative Adversarial Network (GAN) architecture specifically optimized for high-fidelity image super-resolution in digital pathology workflows. The proposed framework enhances standard patch-based architectures by integrating a residual-in-residual dense network (RRDN) back-bone to improve texture feature generation. To prevent checkerboard artifacts and structural blurring, a composite loss function—incorporating pixel-wise mean squared error, perceptual VGG feature distance, and a specialized structural similarity index (SSIM) penalty—guides model training. Validation experiments on open pathology tissue datasets demonstrate that the hybrid network increases image dimensions four-fold while maintaining a Peak Signal-to-Noise Ratio (PSNR) of 31.4 decibels (dB) and an SSIM score of 0.895. Blind evaluations by clinical pathologists confirm that the super-resolved images successfully preserve structural cell boundaries and nuclear chromatin distributions, outperforming standard bicubic interpolation and standard residual network models.

Keywords: Digital pathology, Image super-resolution, Generative adversarial networks, Deep learning, Histopathology, Structural similarity index

Manuscript Timeline: Received: March 18, 2022; Revised: April 25, 2022; Accepted: May 18, 2022; Published: June 1, 2022

Citation: Dupont, J., Laurent, M., & Mercier, P. (2022). A hybrid generative adversarial network for high-fidelity image super-resolution in digital pathology. International Journal of Computer Science and Technology, 3(6), 41–48. DOI: 10.46882/2022/IJCST/000030

Subscribe to read the full article: https://internationalscholarsjournals.org/subscribe-to-read