Purpose: In the rotated treatment such as Tomotherapy, Multileaf Collimator (MLC) movement sequence is a critical component in the radiotherapy (RT) plan and can significantly control patients’ dose distribution. Our study attempted to predict MLC sequence from CTs using deep learning, which can directly get the RT plan.
Methods: This study made the resampling to build the correct MLC sequence and CTs relationship. Equipment of Tomotherapy used in our study has features: when the gantry rotates a round, the gantry will move 8.75mm (pitch*field width, 0.35*2.5cm) in head-foot direction, and 51 phases of control points are recorded. Each control point records 32 pairs (total 32*2 leaves) of MLC locations. After resampling, we got model inputs: 128*128*64 shape CT array (array’s center was the isocenter) with 325*325*8.75mm physical size, and model output: corresponding MLC array with 51*64 shape. Our study enrolled in fifty delivered nasopharyngeal carcinoma (NPC) cases treated by Tomotherapy, 41 for training and 9 for testing. An encode-decode deep learning model ‘MLC-Net’ was built (take model input from 128*128*64 shape to 1*1, then to 51*64), MSE loss was selected.
Results: Mean Square Error (MSE) and Structural Similarity (SSIM) was the validation method for MLC prediction accuracy. The original MLC sequence was stored as an MLC array (shape: 64*1000+, values in [0,1]), and the predicted MLC sequence was also stored as an MLC array with the same format. MSE and SSIM of 9 test cases were 0.023±0.003 and 0.763±0.021.
Conclusion: Our study made a method exploration of directly predicting MLC sequence from CTs in Tomotherapy and predicted MLC has relative ideal MSE and shape similarity. MLC prediction is with the hope to generate auto planning, and which is on beginning stage, more corresponding technical methods and evaluations can be made in the future.
Tomotherapy, Treatment Planning, Optimization
TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation