A Data-Driven Fluence Map Optimization Approach to Mitigate the Risk of Deep Learning Tumor Segmentation Misclassification
R Li1*, N Ebadi2, J Boutilier3, P Rad2, J Buatti1, M De Oliveira1, N Kirby1, N Papanikolaou1, M Bonnen1, A Roy2, (1) University of Texas Health Science Center SA, San Antonio, TX, (2) University of Texas at San Antonio, San Antonio, TX (3) University of Wisconsin-Madison, Madison, WI
Presentations
WE-G-BRC-1 (Wednesday, 7/13/2022) 2:45 PM - 3:45 PM [Eastern Time (GMT-4)]
Ballroom C
Purpose: To incorporate uncertainty and confidence levels of the deep learning (DL) tumor segmentation models into treatment planning, and protect against misclassifications.
Methods: We apply hybrid approaches including Monte Carlo Dropout (MCDO) and deep ensemble to estimate the uncertainty of a spatiotemporal tumor segmentation model. The uncertainty is parameterized via the variance in prediction probabilities and evaluated as a function of the model’s performance (Dice, sensitivity, specificity). Sub-regions are clustered based on uncertainty distributions, and these regions are assigned with different risk (confidence) levels. For example, regions of high variability capture high risk (low confidence). Then, a fluence map optimization model is proposed on the predicted tumor and organ contours, while incorporating the confidence in the predictions. This model contains a probability-based linear-quadratic objective function, multiple conditional value-at-risk (CVaR) constraints on target sub-regions that incorporates confidence levels, and general constraints including max/mean dose on organ-at-risk (OAR) regions. Dosimetry in tumor and OARs is compared between (1) a non-CVaR plan with ground-truth tumor; (2) a non-CVaR plan with predicted tumor and (3) a CVaR-based plan with predicted tumor and uncertainty estimation.
Results: The experiment is conducted on a clinical lung cancer dataset with 16 patients using a 12-field IMRT configuration. The DL model’s performance decreases significantly in the region where uncertainty (variability) is large. Dosimetry analysis shows a significant increase of mean dose on lung, esophagus, and spinal cord when comparing between the non-CVaR plans using predicted tumor (2) versus the ground-truth (1). This increase in dose to surrounding organs can be reduced when employing the CVaR-based plan (3) while maintaining acceptable dose coverage on the ground-truth tumor region.
Conclusion: Our data-driven risk-averse treatment planning approach can reduce the impact of the systematic error/uncertainty produced by deep learning-based segmentation models. We show significant improvements in OAR dose sparing.
Keywords
Taxonomy
TH- External Beam- Photons: IMRT/VMAT dose optimization algorithms
Contact Email