MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Evaluation of a Deep-Learning-Based Seed Identification Model for Prostate Post-Implant Dosimetry: Comparison with Human Operators

Y Yuan*, R Sheu, Y Lo, Mount Sinai Medical Center, New York, NY

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: further evaluate the performance of our previously developed automated seed identification model by comparing with other medical physicists.


Methods: address the challenges of identifying implanted seeds from CT images during post-implant dosimetry (PID), we have developed an automatic algorithm based on a 3D fully convolutional network by modeling the seed localization task as a supervised regression problem that projects the input CT image to a map where each element represents the probability that the corresponding input voxel belongs to a seed. This deep regression model (DRM) significantly suppresses image artifacts and makes the post-processing much easier. Our previous study has demonstrated the proposed model yielded much higher seed detection accuracy than a widely-used commercial seed finder software (VariSeed). In order to compare DRM with human operators, a clinical medical physicist (RS) with 15+ years’ experience on PID retrospectively re-identified implanted seeds on 99 patients in a research environment where he had sufficient time to work on each case. By using his identifications as ground truth, we trained our model with 70 patients and tested it on the rest 29 patients. The accuracy was evaluated by calculating the distance between the predicted seed locations and the ground truth. We also calculated the distance from the original seed locations manually identified by other physicists in clinical use.


Results: took DRM 2 seconds on average to identify seeds on one patient. The median distance between the DRM-identified sees and ground truth was 0.67 mm [25%-75%: 0.37 – 1.17 mm]. At the matching threshold of 3 mm, the DRM correctly identified 94.7% of seeds while the agreement between the clinical identifications and ground truth was 90.8%.


Conclusion: results demonstrate that our DRM can be an efficient and accurate tool to identify seeds on CT images for PID study to improve consistency and efficiency.

Download ePoster [PDF]

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email