Click here to

Session: Therapy General ePoster Viewing [Return to Session]

Improvements On Acute Gastrointestinal Toxicity Modeling by Using Deep-Learning Auto-Segmentation

R Salazar*, J Duryea, A Leone, S Nair, R Mumme, H Baroudi, T Netherton, E Holliday, T Whitaker, K Hoffman, L Court, J Niedzielski, UT MD Anderson Cancer Center, Houston, TX


PO-GePV-T-111 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: To test the hypothesis that, compared with the use of manual contours, the increase in consistency provided by deep learning segmentation will lead to improved NTCP models.

Methods: An auto-segmentation model (nnU-Net) of bowel bag contours was trained and tested on 32 and 8 patients, respectively, using consistently drawn contours approved by a GI radiation oncologist. The bowel bag was then auto-contoured on 195 anal cancer patients CTs, and DVH metrics were extracted on their manually delineated clinical contours (manual-DVHs) as well as those generated automatically (auto-DVHs). Random forest NTCP models for an endpoint of ≥ grade 2, acute, GI tract, CTCAE toxicity were trained with clinicopathological factors and DVH metrics as model predictors. Using the auto-DVHs, 100 models were built following a 10 times 3-fold cross-validation approach on distinct training sets each consisting of 146 randomly selected patients, and then tested on their respective held-out set. To compare to manually derived clinical contours, we repeated this procedure using the manual-DVH metrics. The averages of their respective performance results were compared.

Results: The nnU-Net performance yielded an average surface distance of 2.78±0.88 mm, a Dice score of 0.93±0.02, and a Hausdorff distance of 37±9 mm. The toxicity modeling results with the auto-DVHs showed large improvement over the manual-DVHs. When compared to manual-DVHs, toxicity modeling using the auto-DVH metrics showed an AUC increase from 0.63±0.06 to 0.72±0.05, AUC-PR from 0.55±0.06 to 0.74±0.04, and accuracy from 0.65±0.03 to 0.71±0.04.

Conclusion: These data show that contour delineation methodology can cause variance in calculated DVH metrics, which impacts the toxicity modeling process. Our combined deep learning auto-segmentation and random forest toxicity modeling approach produced superior modeling results when compared to manually derived bowel bag contours.

Funding Support, Disclosures, and Conflict of Interest: Research funded by Varian Medical Systems.


Modeling, NTCP, Segmentation


TH- Dataset Analysis/Biomathematics: Machine learning techniques

Contact Email