Curvature-Sensitive Scheme for Tubular Anatomy Segmentation

Segmenting tubular anatomies from medical images is a difficult task. In addition to all the obstacles typically encountered in the general task of medical image segmentation (obstacles like high inter-subject variability, noise, intra-model differences, etc.), tubular anatomies also have complications which stem from their intrinsic topology. Structures like the airway, aorta, colon, and spine curve in space in unpredictable ways. While this is a challenge in its own right, the problem is perpetuated because this characteristic greatly amplifies the effects of the other difficulties already mentioned (especially anatomical variability).The state-of-the-art in medical image segmentation over the last half decade has been convolutional neural networks. They have dramatically outperformed conventional approaches like multi-atlas segmentation. However, in the context of tubular anatomy segmentation, there is still room for improvement. Our experiments with synthetic data show that the popular convolutional neural network architecture, U-Net struggles to handle the segmentation of long, highly curved tubes, often predicting fragmented segmentations even with considerable training data.In this thesis, we present a new method for handling segmentation of tubular structures that uses both deep learning and conventional techniques in a curvature-sensitive way. Specifically, our method 1) uses a recurrent neural network to unravel tubular anatomies along their centerlines, thereby simplifying the segmentation task and 2) uses classical segmentation technique, region growing in combination with convolutional neural networks to carry out segmentation while maintaining connectivity. We show the effectiveness of our method on synthetic data and on a dataset of high resolution pediatric airway CT scans.Bachelor of Art