"Deep Learning Algorithms for Cardiac Image Classification and Landmark Detection"
With the increase in computational power, deep learning algorithms have become an active field of research over the last 7 years. These data-driven machine learning algorithms have produced good results in many applications, image analysis being one of them. Today, many image analysis tasks in the medical field are done manually, taking valuable time and effort from professionals. Automating some of these tasks could relieve work-load and speed up the healthcare process.
In this thesis, the potential use of deep learning algorithms for medical image analysis will be evaluated. Two problems will be investigated, cardiac image classification and landmark detection in cardiac images.
The algorithms used will be based on two existing deep learning algorithms, the U-net and AlexNet. The deep learning algorithms will be implemented in the deep learning software Caffe, and all training and testing will be ran on an Amazon Web Services instance. The data used for training, testing and validation are provided by the Department of Clinical Physiology at Lund University Hospital. This data is augmented by scaling and rotation to provide a larger and more representative data set for training. The trained algorithm for image classification achieved 98.8% accuracy on validation data, while the algorithm for landmark detection achieved approximately 95% accuracy on validation data.
The image classification algorithms worked well, and it serves as a proof of concept with the potential of being able to solved more clinically difficult problems. With the high accuracy of the landmark detection algorithm largely being due to an imbalance in class distribution, this algorithm, while showing some promise, needs more work to be clinically useful.