Accurate and reproducible semi-automatic liver segmentation using haptic interaction

In this work, we describe and evaluate a semi-automatic method for liver segmentation in CT images using a 3D interface with haptic feedback and stereo graphics. Recently, we reported our fast semi-automatic method using fast marching segmentation. Four users performed initialization of the method for 52 datasets by manually drawing seed-regions directly in 3D using the haptic interface. Here, we evaluate our segmentation method by computing accuracy based on newly obtained manual delineations by two radiologists for 23 datasets. We also show that by performing subsequent segmentation with an interactive deformable model, we can increase segmentation accuracy. Our method shows high reproducibility compared to manual delineation. The mean precision for the manual delineation is 89%, while it is 97% for the fast marching method. With the subsequent deformable mesh segmentation, we obtain a mean precision of 98%. To assess accuracy, we construct a fuzzy ground truth by averaging the manual delineations. The mean sensitivity for the fast marching segmentation is 93% and the specificity is close to 100%. When we apply deformable model segmentation, we obtain a sensitivity increase of three percentage points while the high specificity is maintained. The mean interaction time for the deformable model segmentation is 1.5 minutes.