Room: Room 202
Purpose: To determine the accuracy of a CT-based automated liver segmentation algorithm using a 3D convolutional encoder-decoder network (U-net).
Methods: 21 liver patients datasets (including CT images and peer-reviewed liver contours) are used in this study. Among them, 12 datasets are from liver cancer patients with AJCC staging from I to IVa and rest of the datasets contain patients with normal livers. We trained the U-net, a convolutional encoder-decoder network with skip connections, on 17 datasets. The skip connections allow the decoder portion of the network to predict pixel-precise contour of the segmentation by combining reception fields of different resolutions. We cross-validated our algorithm based on two validation criteria on four validation datasets. First, we computed the intersection-over-union (IoU). The IoU measures the percentage of overlapping volume between the auto-segmentation results and the ground truth segmentation. Second, we measured the surface disparity between the auto-segmentation result and the ground truth in terms of the root mean square error (RMSE), mean absolute error (MAE), maximum absolute error (MaxAE), median of absolute errors (MedianAE), and the 95th percentile Hausdorff distance (HD95).
Results: The deep learning based algorithm demonstrated a noticeable accuracy in automated liver segmentation. In our validation cases, the IoU varied from 80~88%, while RMSE and MAE varied 2.6~5.6 and 1.8~3.3 mm, respectively, which are equivalent to 1~2 voxel distance. Also, from MedianAE, MaxAE, and HD95, which ranged 0.8~2.0, 17~30, and 7~36 mm, it was noticeable that the mean errors are mostly due to a few outlying voxels.
Conclusion: The deep learning based liver segmentation algorithm showed high segmentation accuracy despite that only 17 training dataset had been used for training.