Medical imaging, including computed tomography (CT), magnetic resonance imaging (MRI), mammography, ultrasound, X-ray, and nuclear medicine, is the non-invasive process utilized to create visual representations of interior organs and tissues. Medical imaging’s clinical purpose is to observe health, aid in diagnosis, monitor treatment response, and perform follow-up for disease surveillance. Clinically, interpreting medical images has mostly been performed by human experts such as radiologists or physicians. However, given the wide variety in pathological conditions and the potential fatigue that can result from visual assessment of numerous images, computer-aided diagnosis or detection (CAD) algorithms have been developed and proven to be very helpful. These CAD systems can also provide various functions, such as giving quantitative measurements, extracting radiomics features, and displaying the most important information to assist radiologists’ interpretation. Furthermore, it can even detect suspicious findings and present the malignancy probability using various methods, such as different markers and colors.
The maturity of radiomics analysis with machine learning has provided a very efficient method to build classification models for clinical tasks, including diagnosis, staging, and prognosis prediction. In recent years, neural network methods, a machine learning technique inspired by the human neuronal synapse system, have been widely applied in medical imaging for disease management. The increased volume and quality of digital imaging datasets has created the potential for more accurate and efficient image evaluation using fully automated computer algorithms. However, compared with other machine learning methods such as radiomics, neural networks suffer from several major limitations, including the need for a large dataset to train the deep architecture, the high demand for computing power, and the poor generalization to other datasets not considered in training.
However, during the last 5 years, neural networks have become increasingly popular and have even proven feasible for implementation in clinical practice with the growing availability of big data, enhanced computing power, and novel algorithms. There are many Artificial Intelligence (AI) companies working in this field, and new software being rapidly approved by FDA for clinical use. Deep Learning (DL) algorithms, particularly the convolutional neural networks (CNN), have become the methodology of choice for analyzing medical images. Unlike conventional CAD algorithms, such as radiomics analysis in which task-related features are designed mostly by human experts based on their knowledge about the target domains, deep learning incorporates the feature engineering steps into its learning process. That is, instead of extracting pre-defined features, deep learning only requires pre-processed input data and outcome, discovering its own characteristic information in a self-taught manner. Therefore, the burden of feature engineering has shifted from humans to computers to generate more consistent and reliable outputs.
This thesis will feature radiomics and deep learning-based techniques developed and implemented to extract information from medical images for performing commonly needed clinical tasks, including: lesion detection, organ/tissue segmentation, tumor classification, therapy planning, therapy response prediction, and prognosis prediction.