Master thesis presentation: “Pericardium segmentation in non-contrast cardiac CT images using convolutional neural networks”
On Friday 16 September, 2016, Bolin Shao, MPBME, will be presenting his master thesis with the title “Pericardium segmentation in non-contrast cardiac CT images using convolutional neural networks”
When: 13:00 pm. on Friday, September 16
Where: Lunnerummet (room 3311), Hörsalsvägen 11, 3rd floor
Examiner: Fredrik Kahl
Recent studies shows that the epicardial fat volume is an important indicator for many cardiovascular diseases, such as coronary atherosclerosis. The epicardial fat is the visceral fat that located between the heart and the pericardium. The pericardium structure, i.e. the heart sack, is a thin layer that covers the heart and is barely visible in cardiac CT images.
This thesis proposes a method for automatic pericardium segmentation in non-contrast 3D CT images. The thesis can be divided into two main parts. The first part focuses on pericardium ground truth creation. Given a set of labelled contrast CT images, image registration methods are used to generate an estimated pericardium ground truth for unlabeled non-contrast CT images needed in the second part. Given this estimated labeling, the second part of the thesis concentrates on pericardium segmentation using machine learning algorithms. Convolutional neural networks (CNNs) are trained to classify the voxels in the non-contrast CT images. Multi-atlas techniques are combined with the probability map in order to to define the region of interests. Graph cuts is applied to obtain the final segmentation.
The results shows an average dice coefficient larger than 0.95 on the test images, a comparable number to similar algorithms for contrast CT images.