Deep learning for coronary artery segmentation in CTA images
Year: 2018
Type: dissertation
Abstract: The mortality of cardiovascular diseases (CVD) calls for more knowledge of disease mechanisms and prevention strategies.The SCAPIS study, which is a collaboration between Swedish university hospitals and universities, aims to acquire this knowledge and is the worlds largest study focusing on CVD to date.Among other biomarkers, the SCAPIS study is interested in studying atherosclerotic plaque in coronary arteries which is a major risk factor of CVD.Atherosclerotic plaques can be studied from computed tomography angiography (CTA) images.The first step of the analysis is segmentation of the coronary artery tree, which is tedious and time consuming to do manually.Automatic segmentation techniques can speed up this procedure, but unfortunately, current automatic methods lack the accuracy required and have to be manually corrected.Thus, improving the performance of automatic software could be highly valuable to a large scale study such as SCAPIS.This thesis investigates the viability of a deep learning based automatic segmentation approach and compares its performance to the commercial software Medis QAngioCT Re and to the interobserver variability in manual segmentations from two medical professionals.A version of the state-of-the-art network architecture for vessel wall boundary segmentation, I2I-3D, is implemented in PyTorch and applied to the task.I2I-3D is a fully convolutional network allowing for a variable input size and full volume-to-volume segmentations.It has an encoder-decoder architecture with multiple levels of deep supervision integrated into the decoder.Evaluation on manually corrected segmentations on three CTA images shows that the implemented method outperforms the commercial software for two different segmentation similarity indices; Dice index and modified Hausdorff Distance.However, visual inspection still implies some shortcomings.Although very precise in the segmentation border placement, the network segmentations include broken vessels and unconnected islands of labellings.In contrast, the Medis software's downfall is its inclusion of small, clinically irrelevant vessels.When comparing to the interobserver variability, both automatic approaches perform unfavourably.Nevertheless, I2I-3D shows promising performance despite the relatively limited data set of 25 manually corrected segmentations.The results encourage further development with clear room for improvement.Most importantly, in order to reach full potential the method requires a significantly larger data set.In conclusion, this thesis confirms the viability of using deep learning to segment coronary arteries in cardiac CTA and should be interpreted as a proof-of-concept, rather than the presentation of a finished product.
Author: Fredrik Ring
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2888867878
