Scholarlink

Background/objectives <p>Despite advancements in deep learning and imaging, accurately predicting visual acuity (VA) from colour fundus (CF) images remains difficult. This study aims to evaluate the performance of state-of-the-art convolutional neural networks (CNNs) in predicting VA from wide-field (WF) CF images and assess whether incorporating patient metadata or stratifying by disease can enhance prediction accuracy.</p> Subjects/methods <p>This retrospective study included 3711 patients (17,171 WF-CF images) imaged between 2014 and 2023 at a tertiary ophthalmology centre. Three CNN architectures (ResNet50, EfficientNet-B4, InceptionV3) were used to classify VA into five LogMAR-based categories. Models were trained and tested with and without patient metadata (age, gender) and evaluated across disease-specific subsets (e.g., glaucoma, diabetic retinopathy, cataract). Performance metrics included accuracy, precision, recall, F1 score, and confusion matrix analysis.</p> Results <p>CNN model performance was modest, with mean classification accuracy around 53% and F1 scores ranging from 0.11 to 0.71 across classes. Best performance was observed in the extreme VA categories (Excellent and Bad), with significant misclassification in intermediate classes. Adding metadata and applying disease-specific filtering provided only marginal improvements. Comparison with previous studies reaffirmed that CF-based prediction models consistently underperform relative to those using optical coherence tomography (OCT).</p> Conclusions <p>WF-CF images, even when combined with demographic and diagnostic metadata, do not provide sufficient predictive signal for accurate VA estimation using current CNN methods. These findings underscore fundamental limitations of CF imaging and support the need for OCT or multimodal imaging integration in future VA prediction models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-modal deep learning model for visual acuity prediction from wide field colour fundus imaging

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月3日

查看全文
  • Artem Kim,
  • Assaf Hoogi,
  • Jaime Levy,
  • Leo Joskowicz

摘要

Background/objectives

Despite advancements in deep learning and imaging, accurately predicting visual acuity (VA) from colour fundus (CF) images remains difficult. This study aims to evaluate the performance of state-of-the-art convolutional neural networks (CNNs) in predicting VA from wide-field (WF) CF images and assess whether incorporating patient metadata or stratifying by disease can enhance prediction accuracy.

Subjects/methods

This retrospective study included 3711 patients (17,171 WF-CF images) imaged between 2014 and 2023 at a tertiary ophthalmology centre. Three CNN architectures (ResNet50, EfficientNet-B4, InceptionV3) were used to classify VA into five LogMAR-based categories. Models were trained and tested with and without patient metadata (age, gender) and evaluated across disease-specific subsets (e.g., glaucoma, diabetic retinopathy, cataract). Performance metrics included accuracy, precision, recall, F1 score, and confusion matrix analysis.

Results

CNN model performance was modest, with mean classification accuracy around 53% and F1 scores ranging from 0.11 to 0.71 across classes. Best performance was observed in the extreme VA categories (Excellent and Bad), with significant misclassification in intermediate classes. Adding metadata and applying disease-specific filtering provided only marginal improvements. Comparison with previous studies reaffirmed that CF-based prediction models consistently underperform relative to those using optical coherence tomography (OCT).

Conclusions

WF-CF images, even when combined with demographic and diagnostic metadata, do not provide sufficient predictive signal for accurate VA estimation using current CNN methods. These findings underscore fundamental limitations of CF imaging and support the need for OCT or multimodal imaging integration in future VA prediction models.