HeroHassett/LeNet-MobileNetV2-For-Binary-Classification-of-Infectious-Keratitis

Corneal Opacification is the fifth leading cause of bilateral blindness in the world. The most common cause for Corneal Opacification is infectious keratitis, which can be classified mostly in two categories: bacterial or fungal. The treatments of bacterial keratitis and fungal keratitis are very different; therefore, timely and accurate diagnosis of infection type is crucial. The gold standard of diagnosis is through microscopic examination of culture. This requires speedy access to a lab, which is not always attainable, especially in remote or rural areas or developing countries. I hypothesized that machine learning can use corneal images taken from something as simple as a smartphone camera to differentiate between bacterial and fungal keratitis. I implemented a deep learning LeNet model that could provide an accurate image-based diagnosis. The model is trained through supervised learning on 671 total images with 134 images set aside for validation. The model contains two convolutional layers that connect to the output layer, which has a sigmoid activation function, allowing for a probability of accuracy with every diagnosis. The model utilized Adam optimization and performance was measured using AUC and Accuracy metrics, as well as a binary cross entropy loss function. On the validation dataset, the model achieved an AUC of 0.63 and an accuracy of 0.66 after 50 epochs. The model is able to quickly differentiate between the two infections and has the potential to help people in resource-poor regions and countries gain expedient access to a diagnosis, allowing treatment that could prevent blindness.