This paper follows previous works on the construction of interactive medical image segmentation system, allowing quick volume segmentation requiring minimal intervention of the human operator. This paper contributes to tackle this problem enhancing the previously proposed Active Learning image segmentation system with Domain Knowledge. Active Learning iterates the following process: first, a classifier is trained on the basis of a set of image features extrated for each training labeled voxel; second, a human operator is presented with the most uncertain unlabeled voxels to select some of them for inclusion in the training set assigining corresponding label. Finally, image segmentation is produced by voxel classification of the entire volume with the resulting classifier. The approach has been applied to the segmentation of the thrombus in CTA data of Abdominal Aortic Aneurysm (AAA) patients. The Domain Knowledge referring to the expected shape of the target structures is used to filter out undesired region detections in a post-processing step. We report computational experiments over 6 abdominal CTA datasets consisting. The performance measure is the true positive rate (TPR). Surface rendering provides a 3D visualization of the segmented thrombus. A few Active Learning iterations achieve accurate segmentation in areas where it is difficult to distinguish the anatomical structures due to noise conditions and similarity of gray levels between the thrombus and other structures.
展开▼
