Domain Knowledge Driven 3D Dose Prediction Using Moment-Based Loss Function
G Jhanwar1*, N Dahiya2, P Ghahremani3, M Zarepisheh4, S Nadeem5, (1) Memorial Sloan Kettering Cancer Center, New York, NY, (2) Georgia Institute Of Technology, ,,(3) Memorial Sloan Kettering Cancer Center, ,,(4) Memorial Sloan Kettering, New York, NY, (5) Memorial Sloan Kettering Cancer Center, New York, NY
Presentations
PO-GePV-M-4 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: To introduce a novel moment-based loss function for predicting 3D dose distribution for the challenging conventional lung IMRT plans. The moment-based loss function is convex and differentiable and can easily incorporate clinical dose volume histogram (DVH) domain knowledge in any deep learning framework without computational overhead.
Methods: We used a large dataset of 360 (240 for training, 50 for validation and 70 for testing) challenging conventional lung patients with 2Gy x 30 fractions to train the deep learning (DL) model using clinically treated plans at our institution. We trained a UNet like CNN architecture using computed tomography (CT), planning target volume (PTV) and organ-at-risk contours (OAR) as input to infer corresponding voxel-wise 3D dose distribution. We evaluated three different loss functions: 1) Mean Absolute Error (MAE) Loss 2) MAE Loss + DVH Loss 3) MAE Loss + Moments Loss. The quality of the predictions was compared using different DVH metrics as well as dose-score and DVH-score, recently introduced by the AAPM knowledge-based planning grand challenge.
Results: Model with MAE + Moment loss function outperformed the MAE and MAE+DVH loss with DVH-score of 2.66 ± 1.40 compared to DVH-score of 2.88 ± 1.39 and 2.79 ± 1.52 for the other two respectively. Model with MAE+Moment loss also converged twice as fast as MAE+DVH loss, with training time of ~7 hours compared to ~14 hours for MAE+DVH Loss. Sufficient improvement was found in D95 and D99 dose prediction error for PTV with better predictions for mean/max dose for OARs specially cord and esophagus.
Conclusion: DVH metrics are widely accepted evaluation criteria in the clinic. However, incorporating them into the 3D dose prediction model is challenging due to their non-convexity and non-differentiability. Moments provide a mathematically rigorous and computationally efficient way to incorporate DVH information in any deep learning architecture.
Funding Support, Disclosures, and Conflict of Interest: This work was partially supported by the MSK Cancer Center Support Grant/Core Grant (NIH P30 CA008748)
Keywords
Radiation Therapy, Dose Volume Histograms
Taxonomy
IM/TH- Mathematical/Statistical Foundational Skills: Machine Learning
Contact Email