Quantitative and Qualitative Evaluation of Translating An Adult Auto-Contouring Algorithm to Pediatric Patients
S Hernandez*1,2, L Court1,2, D Rhee1,2, M Soliman2, C Cardenas1,2, (1) UT MD Anderson Cancer Center UTHealth Graduate School of Biomedical Sciences, Houston, TX, (2) The University of Texas MD Anderson Cancer Center, Houston, TX
Presentations
SU-IePD-TRACK 3-1 (Sunday, 7/25/2021) 12:30 PM - 1:00 PM [Eastern Time (GMT-4)]
Purpose: To perform a quantitative and qualitative assessment of using an adult auto-contouring algorithm to generate pediatric normal tissue contours.
Methods: A clinically deployed, adult head-and-neck (HN) auto-contouring model was originally trained to delineate 16 normal tissue structures on 3,495 retrospective, adult HN patients. This work tested the model’s ability to segment the same 16 structures on 143 retrospective pediatric CSI CT scans. Dice similarity coefficient (DSC) and Hausdorff distance (HD) were computed for each of the 16 HN structures. The DSC and HD were compared to the metrics achieved on the adult data set, using an unpaired t-test for each structure (p<0.05 as statistically significant). To assess the clinical acceptability of the contours, we conducted a physician review using 20 patients and a scoring system of 1=no edits, 2=minor edits, and 3=major edits. The clinical acceptability of the pediatric contours was compared to what was achieved on the adult contours.
Results: The average DSC±1σ/HD±1σ(cm) achieved on the brain, brainstem, eyes(avg), lenses(avg), optic nerves(avg), parotids(avg), cochleae(avg), spinal cord, esophagus and optic chiasm, were 0.98±0.01/0.97±1.51, 0.80±0.08/0.97±0.76, 0.87±0.53/0.46±4.4, 0.65±0.12/0.51±4.10, 0.61±0.17/0.37±4.97, 0.60±0.23/1.17±8.10, 0.53±0.17/0.48±0.24, 0.67±0.15/7.40±7.10, 0.57±0.13/3.31±1.40, and 0.34±0.16/0.89±5.40, respectively. The difference between the adult and pediatric DSC values was not statistically significant. The difference between the HD of the esophagus was statistically significant (p=0.0001). Physician review scored 8/16 pediatric and 14/16 adult contours as clinical acceptable (no or minor edits), respectively. Clinically unacceptable pediatric structures included the lenses (5%), optic nerves (5%), optic chiasm (42%), parotids (55%), mandible (5%), and esophagus (11%). The number of structures requiring major edits for patients <2 and >2 years old was 6 and 1, respectively.
Conclusion: Our results suggest that apart from very young patients (<2 years old), adult auto-contouring algorithms can automatically generate clinically acceptable contours on pediatric patients for a majority of HN structures.
ePosters
Keywords
Not Applicable / None Entered.
Taxonomy
Contact Email