| Popis: |
Deep Learning based methods have become immensely popular in recent years. While these methods look very promising, they are not always translated to the medical images domain as-is due to the nature of biomedical images being different from the natural images. The problems arising in the medical images domain are considerably different. In addition, another problem is that the amount of medical images available publicly is rather limited and hence the challenge lies in enabling methods to learn from limited medical data. In this dissertation we worked on different computer vision tasks such as image classification and image segmentation and on a diverse set of medical data such as Computed Tomography (CT) images, Histopathology Whole Slide Image (WSI) and intraoral scans with the recurrent underlying theme as learning from limited data. We begin with proposing novel deep learning based solutions for liver fibrosis and NAS scores classification. Liver fibrosis is a disease that brings in very subtle changes in the liver texture in the different stages of the disease. Hence we design a network suitable for discriminating between subtle changes in texture by focusing on the individual pixel and its neighborhood. We then proceed on to develop a framework aimed at early prognosis of subjects at risk of developing liver cancer. Next, we shift our focus to the problem of tooth segmentation from 3D intraoral scans. We propose a novel deep learning based framework for tooth segmentation from intraoral scans and with a simplified tooth mesh cell representation as the input data. In this dissertation, we also make an interesting observation regarding the data under discussion and explore a relevant question that came up naturally - how much representation can be learnt from a single intraoral scan.Overall, our solutions, approaches and findings together provide new insights into the possibilities and liabilities associated with learning from limited data. We believe which these findings can help develop robust solutions for critical medical applications. |
