Room: ePoster Forums
Purpose: Post-implant dosimetry (PID) is an essential component of prostate brachytherapy. However, identifying implanted seeds from CT images is a challenging task not only because of the severe metal artifacts introduced by the seeds, but also the highly-overlapped appearance when multiple seeds clustered together. As the result, seed identification is a time-consuming procedure that usually takes 10 â€“ 20 minutes. The purpose of this study is to evaluate a fully automatic framework based on deep fully convolutional network (FCN) for Pd-103 seed identification in PID.
Methods: Using the seed locations identified in clinically approved PID plans, our FCN model, which includes 76 layers and 2.3M trainable parameters, was trained to infer the probability of seed locations given sim-CT images. Eighty-six prostate patients who had Pd-103 seeds implant were included in this study, in which 70 patients were used for model training. Our FCN is a fully 3D model that employs convolution and max-pooling to aggregate contextual information of VOI that encompasses the prostate, and uses transpose convolution and skip connections for better determination of seed locations. The predictive accuracy was evaluated by calculating the distance between the predicted seed locations and the locations identified by physicists (ground truth).
Results: The trained model was able to recover the number of implanted seeds on 16 independent testing patients, taking only 2 seconds per patient on average. The median distance between the FCN-identified seeds and ground truth was 0.66 mm [25%-75%: 0.36 â€“ 1.22mm]. A total of 1197 of 1273 seeds (94%) were correctly identified at a matching threshold of 3 mm.
Conclusion: The preliminary results demonstrate that our 3D FCN-based framework has the potential to be an accurate and efficient tool to identify seeds on CT images for PID study. A dosimetric study with larger database will be conducted in the future.
Not Applicable / None Entered.