Learn to recognize pathological myopia in fundus images using bag-of-feature and sparse learning approach

Pathological myopia is a leading cause of visual impairment, and can lead to blindness in children if left undetected. We present a bag-of-feature and sparse learning based framework to automatically recognize pathological myopia in retinal fundus images and discover the most related visual features corresponding to the retinal changes in pathological myopia. In the learning phase, the codebook for the bag-of-feature model and the classification model are first learnt, and the top related visual features are discovered via sparse learning con-currently. In the testing phase, for a given retinal fundus image, local features are first extracted and then quantized with the learned codebook to obtain the global feature. Finally, the classification model is used to determine the presence of pathological myopia. Our results on a population based study dataset of 2258 images achieve a 0.964 ± 0.007 AUC value and 90.6±1.0% balanced accuracy at a 85.0% specificity. The results are promising for further development and validation of this framework.