Purpose: Investigate the feasibility of using neural network object detection algorithms for fiducial tracking during intrafractional motion. Develop a protocol for handling cases where fiducials are not found by the algorithm, or there are false positives.
Methods: Images from a pancreas patient with four implanted fiducials were used. Projections from the pre-treatment half-fan CBCT were manually labelled to identify fiducial positions. 90 equally spaced projections were used as training data and 32 others were used as validation data. The object detection algorithm was Keras RetinaNet, with ResNet50 as the backbone and TensorFlow version 1.12 as the backend. The model was used to detect fiducials in pre-treatment fluoroscopy images obtained with the patient breathing freely. Candidate fiducial positions returned by the RetinaNet algorithm were reduced first by non-max suppression; then a threshold confidence value of 0.12 was set; then any candidate fiducial more than 200 pixels away from any of the others was discarded. The averaged superior-inferior position of the candidate positions with the four highest confidences was calculated and compared to results obtained with a template matching algorithm.
Results: Mean average precision obtained on the validation data set was > 0.9. Fiducials were found with confidence levels reported by RetinaNet of near 0.1 to 0.5 on the fluoroscopy images. The positions determined with RetinaNet were in close agreement with positions determined by template matching, over a breathing cycle with peak to peak amplitude of 10 mm. Several positions determined with the neural network were incorrect, resulting from false positives in the detection stage.
Conclusion: Fiducials can be accurately located using neural net object detection. Ongoing work will optimize detection parameters and robustly determine the centroid position when individual fiducials are not correctly identified.TensorFlow, the TensorFlow logo and any related marks are trademarks of Google Inc.
Funding Support, Disclosures, and Conflict of Interest: Research funded by Varian.