Abstract
Purpose :
Meta-learning provides a promising pathway to train convolutional neural networks (CNNs) under data scarcity, yet its accuracy is restricted in screening strabismus due to the inability to make full use of eye information of data. To alleviate this issue, this study proposed a method by combining a meta-learning approach with image processing methods to fully extract the information that helps determine strabismus, thereby improving the classification accuracy for screening strabismus
Methods :
The meta-learning approach was initially pretrained on a public dataset to obtain a well-generalized embedding network for extracting relevant pixel features while image processing methods were used to extract the position features of eye regions (e.g., iris position, corneal light reflex) as supplementary features. The dimensionality of pixel features was reduced to integrate with the low-dimensional supplementary features via principal component analysis, and the integrated features were used to train a support vector machine classifier for performing strabismus screening. A total of 60 images (30 normal and 30 strabismus) were used to verify the effectiveness of the proposed method, and its classification performance was assessed by computing the accuracy, specificity, and sensitivity through 5000 experiments.
Results :
The classification accuracy using only the meta-learning approach achieved 0.709 with a sensitivity (i.e., correctly classify strabismus) of 0.740 and a specificity (i.e., correctly classify normal) of 0.678 while the proposed method achieved 0.805 classification accuracy with a sensitivity of 0.768 and a specificity of 0.842 in strabismus screening.
Conclusions :
The proposed strabismus screening method achieved promising classification accuracy and improved the classification performance by about 10% than using the meta-learning approach alone under data scarcity. We expect that this combination approach can be an effective solution in medical fields where data scarcity is common.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.