FuNet-40: fundus disease/abnormality classification using ensemble of fine-tuned pretrained convolution models

Age-related macular diseases (AMD) are common reason for visual impairment in humans. These anomalies can result from a variety of illnesses and disorders. Currently, skilled medical professionals make this diagnosis by visually inspecting the pictures. The study of ophthalmology is moving towards the creation of a computer-aided diagnosis (CAD) system for the identification of ocular disorders and diseases. It is crucial to classify these illnesses as accurately and without any false-negatives as possible. However, a person’s retina could be impacted by a number of fundus problems. So, for the purpose of classifying these diseases and pathologies into multiple categories, we study a case with 40 possible class labels. Random Forest was found to be 72.57% accurate based on global features extracted from the dataset and an analysis of the performance of several machine learning models. The dataset was then trained using a few user-defined and pretrained models, and it was found that EfficientNet B1 outperformed all other deep learning models in terms of test accuracy (90.2%), precision (0.993), recall (0.992), F1 score (0.8737), and (=0.2) score. All the models were trained on a set of 1166 images, validated on 250 images, and tested on 250 images.