Session Item

Accurate fiducial marker segmentation is essential for kV-guided intra-fraction motion management to enable stereotactic ablative radiotherapy of the pancreas. We developed a compact convolutional neural network (CNN) model with 4 layers for marker segmentation in the prostate with excellent sensitivity (99.0%) and specificity (98.9%). Deep learning techniques don’t require additional learning imaging, prior marker properties (such as shape or orientation) and they are applicable to kV images. In this study, we further develop our CNN model for marker tracking applied to pancreatic cancer patient data.

We evaluated a CNN with 6 layers and a transfer learning approach from pretrained compact CNN. Training data from the ethics approved SPAN-C Trial for pancreas SABR was utilised. The training dataset contained both cone beam computed tomography (CBCT) projections and kV triggered images acquired during treatment (a total of 23 fractions of 7 patients) for pancreas patients with implanted fiducial markers. Data augmentation was also performed for subimages which contained markers. The total dataset had 1.3 million subimages. The CNN with 6 layers was trained on the full dataset and the transfer learning approach was trained with 32.3% of the full dataset. Cross validation based early stopping was employed to avoid overfitting for both. The performance of each model was tested on unseen CBCT and kV images from 5 fractions of 2 patients. The sensitivity, specificity, the Receiver Operating Characteristic (ROC) curve and the Area Under the ROC (AUC) plot were evaluated. The root-mean-square error (RMSE) was calculated for the centroid of the markers predicted by the CNN models, relative to the manually segmented ground truth.

The sensitivity and specificity of the fully trained CNN was 98.4% and 99.0%, respectively, while the transfer learning model had 94.3% and 99.3%, respectively. The AUC of the fully trained model and that of the transfer learning model was 0.9887 and 0.9889, respectively. The mean RMSE of the fully trained CNN was 0.20 ± 0.03 mm and 0.35 ± 0.05 mm in x and y directions (of kV image), respectively, while the transfer learning had 0.15 ± 0.02 mm and 0.35 ± 0.04 mm in x and y directions, respectively.

A deep learning approach was implemented to classify implanted fiducial markers in pancreatic cancer patient data. The accuracy of marker position prediction by the CNN models from the ground truth was submillimeter as required for stereotactic ablative radiotherapy of the pancreas.