Recently, two undergraduate students mentored by Dr. Hong Jin of Nanchang University have published two research papers on ophthalmic medical image analysis in the international journal Neural Networks (CAS Q2, Top Journal; CCF B in the field of AI). These studies propose innovative solutions addressing the challenge of domain generalization for retinal vessel segmentation and the need for precise detection of retinal diseases, respectively. The related achievements provide practical technical references for the intelligent analysis of ophthalmic medical images.

Progress 1: DGSSA Model Tackles Domain Generalization Challenge in Retinal Vessel Segmentation, Providing a Precise Tool for Vascular Morphology Analysis

This study addresses the "domain generalization" challenge in retinal vessel segmentation. Its core contribution lies in breaking through the limitation of "style enhancement-dominated" approaches in current domain generalization research — existing methods generally adapt to domain shifts only by optimizing image styles (e.g., simple color/contrast changes, or encoding and enhancing decoupled style information), often overlooking structural distribution differences (such as anatomical variations in vessel branching patterns, diameter ratios, density among different populations) that also exist between training data and unseen datasets. To this end, the study proposes for the first time a domain generalization framework (DGSSA) that synergizes "structural enhancement + style enhancement": it generates artificial vessel structures simulating real vascular growth patterns using a spatial colonization algorithm and pairs them with an improved Pix2Pix model to construct diverse pseudo-retinal images, thereby supplementing structural distribution diversity at the source. Simultaneously, it combines Pix2Pix for style enhancement to cover differences in imaging conditions. Validated on four major open-source datasets (DRIVE, CHASEDB1, HRF, STARE) using DeepLabv3+ as the segmentation backbone, the framework achieved an average Dice Similarity Coefficient of 77.98%, with scores of 82.17% on STARE and 72.66% on HRF (which contains many tiny vessels and significant structural differences). It can adapt to images from different sources without fine-tuning on target domain data, providing a more robust tool for clinical vascular morphology analysis.

This research result was published under the title "DGSSA: Domain generalization with structural and stylistic augmentation for retinal vessel segmentation." Nanchang University undergraduate Liu Bo is the first author, Dr. Hong Jin is the corresponding author, and Nanchang University is the primary affiliation.

Link to the publication: https://doi.org/10.1016/j.neunet.2025.108118

Progress 2: WaveNet-SF Model Integrates Spatial-Frequency Domain Learning, Achieving Ultra-Precise Retinal Disease Detection

This research focuses on challenges in optical coherence tomography (OCT) image analysis, such as speckle noise interference, variable lesion morphology, and blurred edge details, proposing the hybrid WaveNet-SF model. Its core contribution is: decomposing OCT images into low-frequency components (preserving global structure for capturing large lesions) and high-frequency components (focusing on edge details for identifying small lesions) via wavelet transform; designing a Multi-Scale Wavelet Spatial Attention (MSW-SA) module to precisely locate different lesions; and introducing a High-Frequency Feature Compensation (HFFC) module to restore edge information and suppress noise, thereby achieving collaborative learning of spatial and frequency domain features. In tests on two major open-source datasets, OCT-C8 (8 disease categories) and OCT2017 (4 disease categories), the model achieved classification accuracies of 97.82% and 99.58%, respectively. Even under high-noise scenarios (PSNR=19.28 dB), accuracy on the OCT2017 dataset remained at 94.78%. Furthermore, without fine-tuning, it achieved 71.84% accuracy on the cross-source OCT-Ext8 external dataset, providing a practical technical solution for automated retinal disease detection.

This research result was published under the title "WaveNet-SF: A hybrid network for retinal disease detection based on wavelet transform in spatial-frequency domain." Nanchang University undergraduate Cheng Jilan is the first author, Dr. Hong Jin is the corresponding author, and Nanchang University is the primary affiliation.

Link to the publication: https://doi.org/10.1016/j.neunet.2025.108189

Both studies were supported by the National Natural Science Foundation of China (Grant No. 62466033) and the Natural Science Foundation of Jiangxi Province (Grant No. 20242BAB20070). The achievements not only address specific technical challenges in retinal medical image analysis but also offer reference for other medical imaging tasks through the proposed concepts of structure-style augmentation and spatial-frequency collaborative learning. Moreover, the completion and publication of high-level research papers with undergraduates as first authors demonstrate the effectiveness of Nanchang University in cultivating undergraduate research and innovation capabilities.