Evaluating High-Resolution Vessel Mask-to-Fundus Translation under Non-Monotonic GAN Dynamics

DOI

10.2991/978-94-6239-648-7_98How to use a DOI?

Keywords

Retinal fundus synthesis; Vessel mask; Conditional GAN; Pix2Pix; Evaluation metrics

Abstract

Clinically viable retinal fundus synthesis from vessel masks requires photorealistic appearance while maintaining anatomical agreement with the input structure. The task is treated as a structure-sensitive, high-resolution (512 × 512) paired translation problem, with a Pix2Pix-style model as a baseline. Evaluation uses distribution metrics (FID, KID) alongside paired measures (LPIPS, MS-SSIM) to separate set-level realism from target-aligned fidelity. Because adversarial training can vary substantially across epochs, several late-stage checkpoints are compared. The checkpoint with the best FID/KID often differs from the one with the best LPIPS/MS-SSIM, so checkpoint choice depends on whether the goal is realism for augmentation or stricter per-sample correspondence. Qualitative inspection is supported by an auxiliary vessel segmenter that visualizes the input mask, generated fundus image, and re-segmented vessels in a single layout. All experiments follow a fixed protocol on the combined DRIVE + CHASE DB1 training set (48 image–mask pairs) as an in-sample reference. Later checkpoints (e.g., epoch 190) show fewer texture/color artifacts and more consistent vessel structure than earlier ones (e.g., epoch 70), indicating that training stage can strongly affect the realism–fidelity balance under limited paired supervision.

Copyright

© 2026 The Author(s)

Open Access

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Download article (PDF)

TY  - CONF
AU  - Jiaqiang Yang
PY  - 2026
DA  - 2026/04/24
TI  - Evaluating High-Resolution Vessel Mask-to-Fundus Translation under Non-Monotonic GAN Dynamics
BT  - Proceedings of the International Workshop on Advances in Deep Learning for Image Analysis and Computer Vision (IWADIC 2025)
PB  - Atlantis Press
SP  - 913
EP  - 922
SN  - 2352-538X
UR  - https://doi.org/10.2991/978-94-6239-648-7_98
DO  - 10.2991/978-94-6239-648-7_98
ID  - Yang2026
ER  -