Recognition of Leukemic Retinopathy Using Knowledge of Diabetic Retinopathy

Leukemic retinopathy is an ocular manifestation associated with leukemia and is considered a key indicator for early diagnosis and treatment of the disease. However, the study of this pathology faces a challenge due to the scarcity of images, making it difficult to develop and train deep learning models for its classification. A dual transfer learning methodology is used to take advantage of the knowledge gained from a well-documented disease such as diabetic retinopathy and apply it to leukemic retinopathy, a less studied disease. First, transfer learning by fine-tuning is implemented on several Convolutional Neural Networks (CNN), including Compact, ResNet-18, AlexNet, RegNet, VGG16, InceptionV3, and EfficientNetV2, using diabetic retinopathy images. Then, the CNN model with the best result for diabetic retinopathy detection is subjected to a second transfer learning process but with a feature-extraction approach using leukemic retinopathy images. The results show that EfficientNetV2 achieved an accuracy of 82.14% in tests using diabetic retinopathy images. Similarly, this model demonstrated an accuracy of 93.33% when testing for leukemic retinopathy. This approach represents a significant advance in identifying leukemic retinopathy, showing the advantages of transfer learning in data-limited scenarios and the benefit of using knowledge of similar diseases to enrich the study of less documented pathologies.