Deep learning methods to aid lesion analysis in medical images

Abstract

In this thesis, novel deep learning-based methods and extensions to existing methods that aid the segmentation and analysis of lesions have been presented. In the first half of this thesis (Chapter 2 and Chapter 3), we studied the role of uncertainty estimation methods in improving the performance of deep learning-based lesion segmentation models. Our work on false-positive reduction (Chapter 2) showed that the miscalibration present in the predictions prevents a meaningful interpretation of uncertainty estimates. We demonstrated the benefits of using more expressive distributions to model the latent space distribution to handle the segmentation of ambiguous images (Chapter 3). In the second half of the thesis (Chapter 4 and Chapter 5), we showed that the use of learned landmark correspondences can improve the registration of organs that exhibit a high degree of motion or deformation, with a focus on the clinical task of lesion co-localization. We extended an existing deep learning-based landmark correspondence prediction model, DCNN-Match, by introducing a soft mask loss term and demonstrated an improvement in lung CT registration in the absence of lung masks (Chapter 4). We also showed that using learned landmark correspondences, on or near vessels in the liver, can improve lesion co-localization over standard intensity-based image registration (Chapter 5).