MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Delay-Line Based Optical Reservoir Computing Towards Real-Time Respiratory Motion Prediction

Z Liang1, C Shi2, X Tang3*, J Li4, X Shen5, Z Huang6, (1) Rensselaer Polytechinic Institute, Troy, NY, (2) Memorial Sloan Kettering Cancer Center, Marlboro, NJ, (3) Yale New Haven Hospital, New Haven, CT, (4) Rensselaer Polytechnic Institute, Troy, NY, (5) Rensselaer Polytechnic Institute, New York, NY, (6) Rensselaer Polytechnic Institute, Troy, NY

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Commonly employed learning (ML) algorithms include multi-layer perceptron neural network, adaptive boosting and multi-layer perceptron neural network, and bi-directional long-short term memory (LSTM) etc. These methods often demand minutes to hours to establish a training model. When irradiating tumors with respiratory induced motion, patient respiration patterns might change momentarily. Any breathing irregularity would result in an unacceptable prediction error from the existing model. In this research, we study software and hardware combined approach, namely optical delay-line reservoir computer to develop a dynamically adjusted model for real-time motion prediction.


Methods: The optical true-time delay reservoir (TDR) is implemented in a benchtop setup. The reservoir consists of discrete optical components in a ring topology similar to an optical oscillator. The optical gain is adjusted to attain a short-term decaying memory. The input layer and output layer will be executed in Matlab or C. The continuously streamed data would feed the training model to adapt to the varying breath pattern. We apply the TDR system to predict 4 patient breathing acquired using RPM.


Results: The regular prediction method involves a training model established on a large data set equivalent to ~ 90 seconds of breath pattern data of a patient. In this research, we have systematically evaluated how the size of the training data may impact the TDR model, therefore, impact the motion prediction accuracy. When a minimum data set is used to establish the model, clinically it would give rise to a significant reduction in inpatient treatment time.


Conclusion: The computation kernel is implemented in a dedicated photonic TDR hardware that enables the ultra-fast computation speed leading to real-time learning model update. It will allow the dynamic tracking of the respiration induced motion. The prediction result is clinically acceptable and much faster than the existing methods.

Keywords

Motion Artifacts, Optimization, Respiration

Taxonomy

TH- RT Interfraction Motion Management: Development (new technology and techniques)

Contact Email