Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2025 Jun 6:13:1608325.
doi: 10.3389/fcell.2025.1608325. eCollection 2025.

MSLI-Net: retinal disease detection network based on multi-segment localization and multi-scale interaction

Zhenjia Qi  1 , Jin Hong  1 , Jilan Cheng  1 , Guoli Long  1 , Hanyu Wang  2 , Siyue Li  3 , Shuangliang Cao  4
Affiliations

MSLI-Net: retinal disease detection network based on multi-segment localization and multi-scale interaction

Zhenjia Qi et al. Front Cell Dev Biol. .

Abstract

Background: The retina plays a critical role in visual perception, yet lesions affecting it can lead to severe and irreversible visual impairment. Consequently, early diagnosis and precise identification of these retinal lesions are essential for slowing disease progression. Optical coherence tomography (OCT) stands out as a pivotal imaging modality in ophthalmology due to its exceptional performance, while the inherent complexity of retinal structures and significant noise interference present substantial challenges for both manual interpretation and AI-assisted diagnosis.

Methods: We propose MSLI-Net, a novel framework built upon the ResNet50 backbone, which enhances the global receptive field via a multi-scale dilation fusion module (MDF) to better capture long-range dependencies. Additionally, a multi-segmented lesion localization module (LLM) is integrated within each branch of a modified feature pyramid network (FPN) to effectively extract critical features while suppressing background noise through parallel branch refinement, and a wavelet subband spatial attention module (WSSA) is designed to significantly improve the model's overall performance in noise suppression by collaboratively processing and exchanging information between the low- and high-frequency subbands extracted through wavelet decomposition.

Results: Experimental evaluation on the OCT-C8 dataset demonstrates that MSLI-Net achieves 96.72% accuracy in retinopathy classification, underscoring its strong discriminative performance and promising potential for clinical application.

Conclusion: This model provides new research ideas for the early diagnosis of retinal diseases and helps drive the development of future high-precision medical imaging-assisted diagnostic systems.

Keywords: lesion localization; multi-scale feature fusion; noise suppression; retinal disease detection; wavelet transform.

PubMed Disclaimer

Conflict of interest statement

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Figures

FIGURE 1
FIGURE 1
Eight categories of OCT images. (A) AMD, (B) CNV, (C) CSR, (D) DME, (E) DR, (F) DRUSEN, (G) MH, (H) NORMAL.
FIGURE 2
FIGURE 2
Overall architecture of MSLI-Net.
FIGURE 3
FIGURE 3
Overall architecture of MDF.
FIGURE 4
FIGURE 4
Overall architecture of LLM.
FIGURE 5
FIGURE 5
(A) Overall architecture of WSSA; (B) Architecture of MSA.
FIGURE 6
FIGURE 6
Accuracy and loss during the training process.
FIGURE 7
FIGURE 7
The confusion matrix of the results, with an accuracy of 96.93%.
FIGURE 8
FIGURE 8
Test dataset after adding different levels of noise.
FIGURE 9
FIGURE 9
Visualization and analysis of MDF, DFE, MSCB and ASPP.
FIGURE 10
FIGURE 10
Visualization and analysis of SE, cropping strategy in AELGNet, and LLM.
See this image and copyright information in PMC

Similar articles

See all similar articles

Cited by

References

    1. Alaba S. Y., Ball J. E. (2024). WCAM: wavelet convolutional attention module SoutheastCon 2024. SoutheastCon 2024, Atlanta, GA, United States, 15-24 March 2024 (IEEE; ), 854–859.
    1. Awais M., Muller H., Meriaudeau F. (2017). Classification of sd-OCT images using a deep learning approach. IEEE International Conference on Signal and Image Processing Applications, 8120661–492. 10.1109/ICSIPA.2017.8120661 - DOI
    1. Bui P.-N., Le D.-T., Bum J., Choo H. (2024). Multi-scale feature enhancement in multi-task learning for medical image analysis. arXiv preprint arXiv:2412.00351. 10.48550/arXiv.2412.00351 - DOI
    1. Burrus C. S., Gopinath R. A., Guo H. (1998). Wavelets and wavelet transforms. houston edition. Houston, TX: rice university, 98.
    1. Cheng J., Long G., Zhang Z., Qi Z., Wang H., Lu L., et al. (2025). WaveNet-SF: a hybrid network for retinal disease detection based on wavelet transform in the spatial-frequency domain. arXiv preprint arXiv:2501.11854. 10.48550/arXiv.2501.11854 - DOI

LinkOut - more resources