Click here to


Are you sure ?

Yes, do it No, cancel

Impact of CT Scanner Acquisition and Reconstruction Methods On Pediatric Organ Autosegmentation Model Generalizability

Philip M. Adamson1*, Petr Jordan1, Vrunda Bhattbhatt1, Taly Gilat Schmidt2, (1) Varian Medical Systems, 3120 Hansen Way, Palo Alto, CA, USA, (2) Department of Biomedical Engineering, Marquette University and Medical College of Wisconsin, Milwaukee, WI


(Thursday, 7/16/2020) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room: Track 5

Purpose: To investigate the generalizability of a fully convolutional network (FCN) pediatric autosegmentation model on CT scans with acquisition and reconstruction methods that are distinct from those in the training dataset.

Methods: Seven organs-at-risk (OARs) were contoured on 180 pediatric CT scans. Organ-specific FCN models were trained on 140 of these patients from Scanner A (LightSpeed VCT, GE Healthcare, Chicago, Illinois, USA) with filtered-back-projection reconstruction. The models were independently evaluated on 20 patients from Scanner A (male:female ratio = 10:10, mean age = 13.35 ± 3.85 years), and 20 patients from Scanner B (Somatom Definition AS+, Siemens Healthineers AG, Erlangen, Germany) with Safire iterative reconstruction with strength 2 (male:female ratio = 10:10, mean age = 13.10 ± 3.87 years). Average symmetric surface distance (ASSD) and Dice similarity coefficient (DSC) were calculated for each group. A Mann-Whitney U test assessed the statistical significance of the difference between model performance on the two test sets.

Results: The mean ASSDs on the Scanner A test set was compared to that of the Scanner B test set for the pancreas (2.58 mm vs 2.78 mm), left kidney (0.66 mm vs 0.84 mm), right kidney (0.66 mm vs 1.33 mm), stomach (2.26 mm vs 2.80 mm), liver (1.00 mm vs 1.12 mm), heart (1.49 mm vs 1.61 mm) and spleen (0.85 mm vs 1.60 mm), respectively. The corresponding average DSCs were (0.74 vs 0.73), (0.97 vs 0.96), (0.97 vs 0.93), (0.86 vs 0.87), (0.95 vs 0.95) and (0.96 vs 0.93). The difference in DSC between the test sets was not statistically significant (p < 0.05) for all OARs.

Conclusion: This work demonstrates that autosegmentation models generalize to CT scanners and reconstruction methods which were not present in the training dataset, crucial to the wide-spread adoption of tools utilizing autosegementation models trained on limited data.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by NIH U01EB023822. Some authors are employees of Varian Medical Systems.


CT, Segmentation


IM/TH- image Segmentation: CT

Contact Email