In this project, we have the opportunity to explore the application of machine learning techniques in the context of a real-world data set. In ophthalmology, fundus screening is an inexpensive and effective way to detect, as early as possible, diseases that may derive into blindness. Early detection of common ocular disorders is complex since few symptoms are visible in the initial stage of illnesses. For example, the first sign of diabetic eye disease are microaneurysms, which are small and hard to detect.
We aim to develop a Deep Learning model able to automatically recognize eye diseases from color fundus images collected from left and right eyes. Particularly, the idea is to implement a specific class of deep learning neural networks, named Convolutional Neural Networks, which historically represented a breakthrough in building models for image classification. The classification problem is framed into the multi-label image classification, since a patient can present more than one disease among the eight categories.
Architecture of the CNN model:
Accuracy and loss function history of the CNN model fit for the balanced training (green) and validation (red) samples:
