Superpixel Classification for Image Segmentation

Zusammenfassung (Abstract) :

Aortic calcification is one of the main causes for aortic valve stenosis which can lead to a limited function of the aortic valve. In a minimal invasive surgery (MIS) a catheter is inserted into the aorta and the aortic valve is replaced by an artificial valve. Although MIS is considered safer than open surgery, it is still a risky operation as the whole environment in which the catheter is applied is not directly visible. Risk prone areas like branches, calcium deposits, aortic aneurysm and aortic valve stenosis need to be avoided. Especially calcium deposit bears the risk of interrupting the oxygen supply if detached by the catheter. The presented approach applies a Support Vector Machine (SVM) to learn the representation of aortic calcification in computer tomography images which are pre-operative data. Two feature vector extraction methods are presented and compared to each other in terms of prediction results and applicability. In the pixel-wise method the feature vectors are extracted per pixel by windowing over the whole image and computing statistical properties of intensity values. Every pixel is then classified separately. In the segmentwise method an over-segmentation algorithm is first applied to the whole image and features are extracted per segment. The properties of the intensity distribution per segment is used as feature by computing a 32-bin histogram of the intensity values within the segment. Consequently, the classification is carried out on the segments instead of pixels resulting in less training examples needed to train the classifier. It is exemplarily shown that the relationship of neighbouring pixels covered by the segmentation can reach similar results compared to the windowing approach covering only a small neighbourhood of a pixel, and that it can even outperform the pixel-wise approach on some false negatives. Additionally, the segment-wise approach reduces the amount of training data by a factor of more than 1 1354 compared with the pixel-wise approach. The CT images were thankfully provided by Herbert De Praetere, MD from Katholieke Universiteit Leuven.