Development of a classifier of photo images of pathologies for an ultra-small data set

The purpose of the work is to create an algorithm and implement it in a software tool for classifying photographic images of pathology of the central region of the human fundus, detected by autofluorescence research, according to 8 types-patterns: normal, minimal changes, focal, spotted, linear, lace-like, reticular, speckled. Methods used machine learning algorithms (convolutional neural networks) and computer vision (histogram methods, perceptual hash algorithms). The main feature of the task is an ultra-small set of unique photoimages with an accurately diagnosed type of pathology (18 pieces). The accuracy of forecasts when solving a problem using a neural network is 12.5%. The accuracy of the predictions of the developed algorithm using a combination of histograms, perceptual hash and one reference photo of the normal state of the fundus is 60% when selecting the classifier parameters from a set of onephoto for one pathology. When using three reference photos, the norm is 85%. The proposed solution can be used in medicine, ophthalmology, photonics and optics of biological tissues, machine learning for both research and educational purposes.