@misc{oai:repo.qst.go.jp:00084069,
 author = {Ryusuke, Hirai and Yukinobu, Sakata and Shinichiro, Mori and Ryusuke, Hirai and Yukinobu, Sakata and Shinichiro, Mori},
 month = {Oct},
 note = {Purpose/Objective(s)

To achieve high treatment accuracy in respiratory-gated radiotherapy, it is important to obtain a mobile tumor position in 3D space. Instead of tracking a tumor, Mylonas et al. proposed the method of detecting linear fiducial markers in fluoroscopic images with convolutional neural network (CNN) models (Mylonas et al. Med Phys. 2019; 46 2286-97). However, it is difficult to prepare the many marker images required for training the model. Therefore, we propose a method to train a CNN model with 3D simulated training marker images. Furthermore, the proposed model tracks the markers in a pair of orthogonal fluoroscopic images. Owing to no need of taking fluoroscopic images of implanted markers, our method can build the model for the arbitrary-shaped marker more efficiently.

Materials/Methods

To make training images, we cropped patch images, whose size was 31x31x2 pixels, from the position such that 3D points were projected onto the paired fluoroscopic images of non-implanted patients. These patch images were labeled “not-marker images”. The simulated “marker images” were generated by drawing the shape of the marker, which was virtually put at the 3D points with any posture, projected onto the patch images. We built a CNN model by training 1000 marker and not-marker images. The model consisted of four convolutional layers with two max-pooling layers and lastly a fully connected layer to output the marker likelihood. Each convolutional layer has a batch normalization layer and a rectified linear unit. For tracking the marker, the trained model calculated the likelihood of each 3D point by inputting the cropped image from where the 3D point was projected onto the fluoroscopic images and the marker position in 3D space was determined with the likelihood-weighted average of the positions. To reduce the number of calculated 3D points, we used a particle filter. The method was evaluated using three liver fluoroscopic image datasets with an implanted linear fiducial marker 0.5 mm in diameter by 3.0 mm in length. Each dataset consisted of at least one respiratory phase. The positional error of each marker was calculated between the tracked position and the ground truth position in 3D space. In addition, the calculation time was measured on a workstation (Intel Xeon® CPU @ 3.6 GHz).

Results

The mean positional error for three patients was measured to be 0.32 mm with a standard deviation of 0.17 mm. The worst positional error was 1.1 mm. For the first, second and third patient, the mean errors were 0.38, 0.26 and 0.32 mm with a standard deviation of 0.16, 0.13 and 0.16 mm, respectively. The calculation time was within 50.4 milliseconds per frame.

Conclusion

We presented real-time linear fiducial marker tracking for respiratory-gated radiotherapy with a DNN framework. We performed experiments involving three live patients with our proposed method and obtained high accuracy and short computation time., The American Society for Radiation Oncology (ASTRO) Annual meeting 2020},
 title = {Real-time Linear Fiducial Marker Tracking in Respiratory-gated Radiotherapy With a Deep Neural Network},
 year = {2020}
}