RESUMEN
PURPOSE: Considering the scarcity of normal and strabismic images, this study proposed a method that combines a meta-learning approach with image processing methods to improve the classification accuracy when meta-learning alone is used for screening strabismus. METHODS: The meta-learning approach was first pre-trained on a public dataset to obtain a well-generalized embedding network to extract distinctive features of images. On the other hand, the image processing methods were used to extract the position features of eye regions (e.g., iris position, corneal light reflex) as supplementary features to the distinctive features. Afterward, principal component analysis was applied to reduce the dimensionality of distinctive features for integration with low-dimensional supplementary features. The integrated features were then used to train a support vector machine classifier for performing strabismus screening. Sixty 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 5,000 experiments. RESULTS: The proposed method achieved a classification accuracy of 0.805 with a sensitivity (correct classification of strabismus) of 0.768 and a specificity (correct classification of normal) of 0.842, whereas the classification accuracy of using meta-learning alone was 0.709 with a sensitivity of 0.740 and a specificity of 0.678. CONCLUSION: The proposed strabismus screening method achieved promising classification accuracy and gained significant accuracy improvement over using meta-learning alone under data scarcity.
Asunto(s)
Procesamiento de Imagen Asistido por Computador , Estrabismo , Humanos , Procesamiento de Imagen Asistido por Computador/métodos , Aprendizaje Automático , Análisis de Componente Principal , Estrabismo/diagnóstico por imagen , Máquina de Vectores de SoporteRESUMEN
PURPOSE: This study aims to provide an automatic strabismus screening method for people who live in remote areas with poor medical accessibility. MATERIALS AND METHODS: The proposed method first utilizes a pretrained convolutional neural network-based face-detection model and a detector for 68 facial landmarks to extract the eye region for a frontal facial image. Second, Otsu's binarization and the HSV color model are applied to the image to eliminate the influence of eyelashes and canthi. Then, the method samples all of the pixel points on the limbus and applies the least square method to obtain the coordinate of the pupil center. Lastly, we calculated the distances from the pupil center to the medial and lateral canthus to measure the deviation of the positional similarity of two eyes for strabismus screening. RESULT: We used a total of 60 frontal facial images (30 strabismus images, 30 normal images) to validate the proposed method. The average value of the iris positional similarity of normal images was smaller than one of the strabismus images via the method (p-value<0.001). The sample mean and sample standard deviation of the positional similarity of the normal and strabismus images were 1.073 ± 0.014 and 0.039, as well as 1.924 ± 0.169 and 0.472, respectively. CONCLUSION: The experimental results of 60 images show that the proposed method is a promising automatic strabismus screening method for people living in remote areas with poor medical accessibility.