Related Experiment Video
Updated: May 14, 2026
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Multiscale Attention Unet: An Innovative Approach for Segmentation of Optic Disc and Optic Cup in Early Detection of
Yichen Xiao1,2,3,4, Qinghong Lin1,2,3,4,5, Ziwei Xiang1,6
1Eye Institute and Department of Ophthalmology, Eye and ENT Hospital, Fudan University, Shanghai, China.
Objective:
To develop an accurate optic disc (OD) and optic cup (OC) segmentation model for early glaucoma detection and differentiation from high myopia through precise cup-to-disc ratio (CDR) calculation. Timely intervention and differential diagnosis between high myopia and glaucoma are crucial for disease progression and treatment effectiveness.
Design:
Retrospective cross-sectional study of 3 independent ophthalmic data sets.
Subjects:
2814 fundus images from 3 public data sets:ORIGA (650 images); REFUGE (1200 images); G1020 (964 images).
Methods:
Proposed Multiscale Attention Unet (MAUnet) integrates: (1) Incorporating multiscale convolutional (3 × 3 and 5 × 5 kernels) for hierarchical feature extraction. (2) Efficient Channel Attention modules integrate the Efficient Channel Attention Network into the Unet architecture to enhance region-of-interest localization. (3) Hybrid loss function combining Dice loss (macro-structures) and cross-entropy loss (microdetails, λ = 0.5) to solve class imbalance and boundary ambiguity.
Main Outcome Measures:
Dice Similarity Coefficient (DSC) for OD/OC; Intersection over Union; Absolute CDR error; Boundary F1-score.
Results:
Multiscale Attention Unet demonstrated superior segmentation performance across all benchmark data sets. On the ORIGA data set, it achieved DSCs of 0.958 for OD and 0.912 for OC, with an absolute CDR error of 0.038. For the REFUGE data set, the model attained DSC values of 0.972 (OD) and 0.931 (OC), alongside a CDR error of 0.026. Evaluation on the G1020 data set yielded DSCs of 0.963 (OD) and 0.906 (OC), with a CDR error of 0.041. Crucially, MAUnet significantly outperformed comparator models in OC segmentation, evidenced by a mean Boundary F1-score of 0.887 versus the comparator range of 0.812 to 0.852 (P < 0.001), highlighting its exceptional precision for detection and further diagnosis.
Conclusions:
The proposed MAUnet model shows superior performance in OD and OC segmentation, proving to be a clinically reliable tool for early glaucoma detection and monitoring. Its multiscale attention mechanism and balanced loss function enable accurate segmentation, particularly for challenging edge and detail detection, improving glaucoma screening and early diagnosis across diverse populations.
Financial Disclosures:
Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.