Low-shot machine learning for medical image classification

AI and Deep Learning have seen many exciting real-world applications implemented today. The application focus for this project is on automatic medical image classification. Conventional deep learning requires huge amounts of data to be trained on to achieve high performance values. As such, this project aims to develop strategies and models that can produce satisfactory accuracy in classifying medical images given a very small sample size. The performance will be evaluated on the task of classifying normal retinas against diabetic retinopathy retinas and normal lungs against pneumonia infected lungs. First the effects of low-shot training were studied in greater detail by iteratively training a basic convolution neural network model (CNN) with the sample size scaled down each time. The findings showed deteriorating accuracy performance for both tasks. However, for the same model, one task suggested underfitting and the other, overfitting. Three strategies were explored namely, Generative Adversarial Networks (GAN), Transfer Learning and Model Optimization. GAN has the ability to generate synthetic copies of real images. However, in this project our GAN was proven to be unfeasible. Transfer learning relies on previously trained models that are publicly available to use. Model optimization refers to the process of fine-tuning and adjusting the construct of the model to improve its performance. For this project we found that transfer learning and model optimization were successful in tackling our low-shot training problems. For future work and contribution, we hope the strategies developed in this project will be helpful in overcoming similar problem statements. These strategies have the potential to be viable and remarkable solutions to many deep learning challenges faced today.Bachelor of Engineering (Electrical and Electronic Engineering