Purpose: To investigate the feasibility of using deep learning to improve markless lung tumor detection on x-ray radiograph images during real-time motion tracking and correction (RMTC).
Methods: A convolutional neural network (CNN), designed with Matlab (Mathworks, R2020a), was trained, validated and tested with 194 radiograph images acquired at different gantry angles and couch positions from 12 lung cancer patients treated with RMTC during helical tomotherapy. The dataset was divided into training (70%), validation (15%) and test (15%) subsets. The ground truth lung tumor contours on the radiographs were obtained by transferring the contours from DRR with manual editing. All radiographs were normalized based on the intensities of voxels enclosed in a region of interest (ROI) corresponding to the tumor motion ranges on the DRR. Performance of the trained and validated CNN was measured by comparing the predicted target contours of the test data subset with the ground truth using Dice similarity coefficient (DSC) and the distance between the centers from the predicted and the ground truth contours, (dCNN-GT), which was also compared to the distance between the detected center in the RMTC system and the ground truth center (dRMTC-GT).
Results: The training and validation achieved sufficient accuracy (99.68%). The average DSC was 0.76±0.24. The average dCNN-GT was 1.48±1.05 mm, slightly smaller than the average of dRMTC-GT (1.69±1.00 mm). On average, the time needed for a prediction was 119±29 ms with an Intel Core™ i7-6700 CPU@3.40 GHz PC.
Conclusion: The CNN model can quickly predict target center on x-ray radiograph with accuracy equal or slightly better than the current system detection. With further development using large datasets, the deep learning may be used to improve target prediction for real-time markless lung tumor tracking.