Evaluation of deep learning based implanted fiducial markers tracking in pancreatic cancer patients

Real-time target position verification during pancreas stereotactic body radiation therapy (SBRT) is important for the detection of unplanned tumour motions. Fast and accurate fiducial marker segmentation is a Requirement of real-time marker-based verification. Deep learning (DL) segmentation techniques are ideal because they don't require additional learning imaging or prior marker information (e.g., shape, orientation). In this study, we evaluated three DL frameworks for marker tracking applied to pancreatic cancer patient data. The DL frameworks evaluated were (1) a convolutional neural network (CNN) classifier with sliding window, (2) a pretrained you-only-look-once (YOLO) version-4 architecture, and (3) a hybrid CNN-YOLO. Intrafraction kV images collected during pancreas SBRT treatments were used as training data (44 fractions, 2017 frames). All patients had 1-4 implanted fiducial markers. Each model was evaluated on unseen kV images (42 fractions, 2517 frames). The ground truth was calculated from manual segmentation and triangulation of markers in orthogonal paired kV/MV images. The sensitivity, specificity, and area under the precision-recall curve (AUC) were calculated. In addition, the mean-absolute-error (MAE), root-mean-square-error (RMSE) and standard-error-of-mean (SEM) were calculated for the centroid of the markers predicted by the models, relative to the ground truth. The sensitivity and specificity of the CNN model were 99.41% and 99.69%, respectively. The AUC was 0.9998. The average precision of the YOLO model for different values of recall was 96.49%. The MAE of the three models in the left-right, superior-inferior, and anterior-posterior directions were under 0.88 ± 0.11 mm, and the RMSE were under 1.09 ± 0.12 mm. The detection times per frame on a GPU were 48.3, 22.9, and 17.1 milliseconds for the CNN, YOLO, and CNN-YOLO, respectively. The results demonstrate submillimeter accuracy of marker position predicted by DL models compared to the ground truth. The marker detection time was fast enough to meet the requirements for real-time application.