Poster Presentation at the European Congress of Radiology, Vienna, 2019
We discuss a novel method to semi-automatically segment liver parenchyma and vasculature using deep learning and cloud-based tools.
Methods and Materials
We compared the time taken to segment liver parenchyma between 1) Philips Intellispace, an FDA approved liver segmentation and volumetry software and 2) PredibleLiver, a minimal web-based viewer for medical images and segmentation, with basic editing tools. On the Philips Intellispace software, semi-automated segmentation tools like brush, region-grow, multi-ROI interpolation were used to perform the segmentation. On PredibleLiver, only 3D-brush was used to correct the deep learning initialized segmentations. Time taken to annotate parenchyma was measured (in seconds) for 30 triphasic abdomen CT images for 3 technicians with adequate training on both software. To measure the similarity between outputs, DICE score and volume outputs were calculated with the Manually segmented studies using intellispace as ground truth and correlations and summary statistics were obtained.
The deep learning model was trained on 200 tri-phasic CT scans, using a UNET architecture. The model had a DICE of 0.95 on a test dataset containing 70 venous phase abdomen CTs, with a standard deviation of 0.08, and predicted volumes had a correlation of 0.94 (p < 0.05) with volumes of ground-truth segmentations. Median person time involved in completing the segmentation manually for Intellispace was 806.5 s, while it was 204.5 s for the cloud-based semi-automated viewer. Meantime for segmentation decreased by 68% (SD 24%).
This Cloud-based deep learning based segmentation system is more accurate and faster than the workstation based segmentation software.