Related Experiment Video
Updated: Jul 6, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
DeepGraFT: A novel semantic segmentation auxiliary ROI-based deep learning framework for effective fundus
Yinghao Yao1, Jiaying Yang1, Haojun Sun1
1Oujiang Laboratory (Zhejiang Lab for Regenerative Medicine, Vision and Brain Health), Eye Hospital, Wenzhou Medical University, Wenzhou, 325011, Zhejiang, China; National Engineering Research Center of Ophthalmology and Optometry, Eye Hospital, Wenzhou Medical University, Wenzhou, 325027, Zhejiang, China.
An automated system called DeepGraFT accurately grades fundus tessellation (FT) in eye photos, aiding in the prediction of myopia progression and visual impairment. This deep learning tool improves upon existing methods for clinical decision support.
Area of Science:
- Ophthalmology and Medical Imaging
- Artificial Intelligence in Healthcare
- Computational Pathology
Background:
- Fundus tessellation (FT) is a key indicator in myopia, linked to potentially irreversible myopic maculopathy.
- Accurate FT classification in fundus photos is crucial for predicting disease progression and prognosis.
- Current methods for FT detection and classification lack precision, creating a significant unmet clinical need.
Purpose of the Study:
- To introduce DeepGraFT, an automated deep learning system for grading fundus tessellation (FT).
- To enhance the accuracy of FT classification by combining classification and segmentation models.
- To evaluate DeepGraFT's performance on both in-house and independent public datasets.
Main Methods:
- Developed a deep learning model integrating classification (ConvNeXt with transfer learning) and segmentation.
- Utilized a region of interest based on the ETDRS grading system to optimize classification.
- Trained and validated DeepGraFT using the in-house MAGIC cohort and the public UK Biobank cohort.
Main Results:
- DeepGraFT achieved high accuracy in validation: 86.85% (in-house) and 81.50% (UK Biobank).
- The system outperformed traditional machine learning models, showing a 5.57% increase in accuracy.
- Ablation studies confirmed module enhancements, improving accuracy from 79.85% to 86.85%.
- Analysis revealed a significant negative association between FT and spherical equivalent (SE) in the UK Biobank cohort.
Conclusions:
- DeepGraFT demonstrates the significant potential of deep learning for automated FT grading.
- The system can serve as a valuable clinical-decision support tool for predicting pathological myopia progression.
- Automating FT grading can improve the management and prognosis of myopia-related visual impairment.
