Active Learning enhanced with Expert Knowledge for Computed Tomography Image Segmentation

Abstract

Our objective is to create an interactive image segmentation system of theabdominal area allowing quick volume segmentation requiring minimal interventionof the human operator.Our contribution to tackle this problem is to enhance an Active Learning imagesegmentation system with Expert Knowledge, allowing quick and accurate volumesegmentation requiring minimal intervention of the human operator. As a first step,image segmentation is produced by a Random Forest (RF) classifier applied on aset of standard image features. The human operator is presented with the mostuncertain unlabeled voxels to select some of them for inclusion in the training set,retraining the RF classifier. The approach is applied to the segmentation of thethrombus in CTA data of Abdominal Aortic Aneurysm (AAA) patients. The expertknowledge on the expected shape of the target structures is used to filter outundesired detections.We have performed computational experiments over 8 datasets between 216and 560 slices each that consists in real human contrast-enhanced datasets of theabdominal area. The performance measure of the experiments is the true positiverate (TPR). 3-fold cross validation is applied. We average the TPR obtained oneach slice at each iteration of the process, with a corresponding variance value.Surface rendering is computed to show a 3D visualization of the segmentedthrombus.Accurate segmentation is obtained after a few iterations in areas where it isdifficult to distinguish the anatomical structures from surrounding tissues due to avariety of noise conditions and similar the gray levels (i.e. thrombus).