Fig. 4.

Manually delineated ground truth thalamus mask (left column) and the masks automatically generated with each of the three software tools for two cases from the ISBR dataset (first row: case 6, second row: case 18). Case 6 had the highest Hausdorff distance between the ground truth and the 3D-CNN (7.28mm), and case 18 the lowest (3.46 mm). True positive voxels of the automatic masks are shown in blue, false positive voxels in green, and false negative voxels in red. The examples illustrate the observed overestimation of the thalamus volume relative to the manual ground truth by all automatic methods. The examples also indicate that the automatic methods tend to generate smoother segmentation surfaces than manual segmentation