Q Lyu*, M Cao, K Sheng, UCLA School of Medicine, Los Angeles, CA

• 
  Q Lyu

SU-H300-IePD-F4-3 (Sunday, 7/10/2022) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 4

Purpose: Advancing the multi-leaf collimator (MLC) to achieve higher resolution for ultra-small and complex target treatments has reached its physical and engineering limit. In contrast, the sparse orthogonal collimators (SOC) can achieve arbitrary rectangular apertures with substantially improved modulation resolution at the cost of optimization complexity of aperture selection. This study investigates the feasibility of a Learned Iterative Shrinkage-Thresholding Algorithm (LISTA) to improve computation efficiency for SOC planning.

Methods: The sparse rectangular aperture optimization (SRAO) framework was formulated as a least-square dose fidelity objective for dosimetric quality and an L1-norm sparsity term for rectangle selection. A classical iterative solver of the SRAO framework, the proximal gradient descent (PGD) algorithm, was unrolled as a trainable network LISTA. The LISTA network includes 150 phases, each phase containing 3 layers: the gradient descent layer for the data fidelity term, the soft thresholding layer for the L1-norm sparsity term, and the Rectified Linear Unit (ReLU) layer for the non-negative constraint. The network was trained with 40 7-beam coplanar Intensity-modulated radiation therapy (IMRT) plans of one mouse, with a fluence map consisting of multiple deliverable rectangular apertures. LISTA was compared with PGD on convergence speed and dosimetry on a testing plan.

Results: Both LISTA and PGD generated deliverable rectangular apertures. PGD has constant step size and thresholding parameters, while LISTA learned variable parameters at each phase. LISTA achieved lower objective values than PGD with the same number of iterations. The dosimetry of LISTA at 150 iterations is comparable to PGD at 1000 iterations, achieving similar planning target volume (PTV) coverage and normal tissue doses within 3% differences. The optimization time was reduced from 2 minutes to less than 20 seconds.

Conclusion: The trainable network LISTA improved the computation efficiency compared with the classical convex optimization algorithm PGD.

Funding Support, Disclosures, and Conflict of Interest: This research is supported by the AAPM seed grant, DOE Grants Nos. DE-SC0017057 and DE-SC0017687, NIH Grants Nos. R01CA188300, R43CA183390, and R44CA183390.

Keywords

Treatment Planning, Inverse Planning

Taxonomy

TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation

Contact Email