Program Information
Simultaneous Optimization of Surface Surrogates and Inference Model to Estimate Internal Tumor Displacement with Structured Sparsity Regularization
D Ruan1*, M Tajdini2 , A Sawant2 , (1) UCLA School of Medicine, Los Angeles, CA, (2) University of Maryland School of Medicine, Baltimore, Maryland
Presentations
SU-I-GPD-J-80 (Sunday, July 30, 2017) 3:00 PM - 6:00 PM Room: Exhibit Hall
Purpose: Monitoring internal anatomy, including but not limited to tumor is important to accurate dose accumulation and motion-adaptive treatment. Surface monitoring offers safe and efficient external surrogate. This project aims to develop a systematic approach that simultaneously (1) selects the subset of the surface points as reliable surrogates and (2) provides a stable inference model to generate internal displacement estimates.
Methods: We formulate the problem as regularized regression optimization, with linear and nonlinear regression as two inference model settings. In both cases, a structured sparsity regularization is introduced to impose that the same subset of surrogate is used for inference across time. In a pilot study, we extracted surface reading and internal tumor displacement from 4DCT data with 10 phases from 9 patients. Surface displacement was obtained by registration between phases using Elastix. The model, including both the surrogate location and the inference relationship, was trained in a leave-one-out fashion, using a subset of 8 for training and the remaining patient for testing. The inference performance was evaluated by comparing the inferred location to the extracted phase-specific ground-truth. Error statistics are reported in root-mean-squared-erorr (RMSE), mean-absolute-error (MAE) and their variations to assess stability.
Results: We achieved sub-mm accuracy for inferring the internal tumor center. The estimates on tumor boundaries had a higher uncertainty, possibly due to registration errors and patient-specific uncertainty. Points on the chest and abdomen were selected by the algorithm automatically as salient features, agreeing with the observation that thoracic tumors were affected primarily by respiration.
Conclusion: Structured sparsity serves as an automatic feature selection tool in identifying the most informative and consistent external surface point as salient surrogate points. The clustering pattern could be a result of jointly-sparse basis pursuit model, and maybe alleviated with either a nonconvex regularizer or a spatial uniformity term.
Funding Support, Disclosures, and Conflict of Interest: This work is supported in part by NIH 5R01CA169102-05
Contact Email: