Proposal of a Computational Vision Model for the Pre-diagnosis of Anemia Based on the Image of the Ocular Conjunctiva

Anemia is a persistent public health problem in Peru, with significant repercussions on individual quality of life and socio-economic progress at the national level. Although the hemogram is considered the reference method for diagnosing anemia, its need for time and a laboratory setting to be performed can represent a considerable limitation, especially in remote or less developed areas. The aim of the present research is to implement a computer vision model for the pre-diagnosis of anemia from the image of the ocular conjunctiva. The applied methodology was based on 5 phases: obtaining dataset, preprocessing, modeling and classification, feature extraction and implementation of CNN architectures. Several models were run with the classifiers: SVM, RF, MLP, and RNN and feature extraction techniques: SIFT, SURF, ORB and HOG. The model with RF and HOG extractor obtained the highest accuracy with 79%. Finally, deep learning models were explored, adjusting parameters such as the number of neurons, epochs, and samples in each model. Although initially the custom model obtained the highest accuracy of 93.18%, the Inception-ResNet-v2 model, supported by existing studies, was finally chosen and demonstrated a robust accuracy of 90.69% and loss of 0.2788, which is better than the loss of 0.3563 obtained with the comparison model.