Deep Neural Networks for Medical Image Segmentation And Quantitative MR Imaging

Abstract

Neural networks play a valuable role in medical imaging, with applications to both image segmentation (delineation of boundaries) and reconstruction. Traditionally, segmentation is performed manually by a radiologist, but this requires significant time and expertise. Convolutional Neural Networks (CNNs) can efficiently segment some organs with accuracy similar to experts. However, prior to their widespread adoption in medical practice, challenges relating to their performance on data of varying quality must be addressed. These issues are investigated here, with segmentations of MR images produced for use in two clinical studies: one involving non-alcoholic fatty liver disease (NAFLD), and another prostate cancer. Liver volumetry measurements obtained from segmentations are necessary to investigate the pathology of NAFLD. This work focuses on identifying which types of MR images, with differing resolution and abundance of artifacts, yield the most accurate results. Segmenting the prostate into distinct zones provides valuable guidance for an algorithm to detect and diagnose prostate cancer, as the prevalence and presentation of cancer varies by region. Additionally, this work investigates the use of CNNs for reconstruction of quantitative T2 maps of the prostate, which may improve grading of cancer. Reconstruction of T2 maps with traditional methods requires acquisition of multiple images, leading to long acquisition times and limited clinical use. CNNs have the potential to enable reconstruction of T2 maps from fewer images, allowing for more efficient acquisition. Application of CNNs to both image reconstruction and segmentation demonstrate their versatility and great potential for use in biomedical research and medical practice. This work investigates challenges that must be addressed in this process.