A new meta-learning approach effectively detects retinal vein occlusion (RVO) using multimodal imaging, requiring only a few training samples, according to a study that demonstrated that models utilizing CLIP-based prototypical networks outperform traditional methods, achieving high diagnostic accuracy and AUCROC scores even when trained on limited datasets.
This cross-sectional study utilized a meta-learning framework, leveraging a training dataset of 1,254 color fundus images representing 39 distinct fundus diseases.
The study’s methodology included 2 meta-testing datasets:
- A publicly available dataset
- An independent set from Kandze Prefecture People’s Hospital.
The proposed models featured 2 primary components:
- Feature extraction networks
- Prototypical networks (PNs),
The models utilized ResNet and Contrastive Language-Image Pre-Training (CLIP) for the feature extraction process. Performance metrics such as accuracy, area under the receiver operating characteristic curve (AUCROC), F1-score, and recall were employed to evaluate the algorithms.
CLIP-based PNs consistently outperformed competitors across all meta-testing datasets. For the public APTOS dataset, the meta-learning algorithms achieved an accuracy of 86.06% and an AUCROC of 0.87 with just 16 training images. In datasets from the hospital, the algorithms demonstrated remarkable diagnostic capabilities, achieving AUCROC values exceeding 0.99 with as few as 4 training images. The models also excelled in detecting RVO from fluorescein angiography images, with AUCROC scores above 0.93, despite the training dataset lacking these specific images.
Reference
Jiachu D, Luo L, Xie M, et al. A Meta-Learning Approach for Classifying Multimodal Retinal Images of Retinal Vein Occlusion With Limited Data. Transl Vis Sci Technol. 2024;13(9):22. doi: 10.1167/tvst.13.9.22. PMID: 39297809; PMCID: PMC11421671.
Related Content
