MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Surface-Driven MRI-US Registration Using Weakly-Supervised Learning in Prostate Brachytherapy

Q Zeng , J Jeong , Y Lei , Z Tian , T Wang , X Dong , A Jani , P Patel , H Mao , W Curran , T Liu , X Yang*, Emory Univ, Atlanta, GA

Presentations

(Monday, 7/15/2019) 7:30 AM - 9:30 AM

Room: 303

Purpose: To develop a deformable MRI-ultrasound (US) registration method based on weakly-supervised deep-learning to integrate multiparametric MRI into real-time US imaging for tumor-targeted catheter placement in high-dose-rate (HDR) prostate brachytherapy.

Methods: We propose a weakly-supervised, surface-driven, patch-based MRI-US registration framework to learn 3D voxel-wise correspondence from higher-level label correspondence thereby avoiding conventional intensity-based image similarity measures. During training, a convolutional neural network was optimized by generating a dense deformation field (DDF) that deformed a set of available anatomical labels from the MR images to their corresponding US counterparts. These prostate surface label pairs are required during training and can be spatially aligned by minimizing a logistic loss function of the deformed MRI prostate and the US prostate. During deformation prediction, image patch pairs extracted from new MRI and US images were fed into the well-trained network to predict an optimal DDF, realizing a fully automatic, label-free, real-time and deformable registration. Our approach was assessed quantitatively by calculating the centroid distance of post-registration prostate and was also compared to the whole-image-wise surface-driven (WSD) method.

Results: We conducted a clinical study to evaluate our proposed registration method with 25 prostate patients treated with HDR brachytherapy. The Dice similarity coefficient between deformed MRI and US prostates was 0.78 for the WSD method and 0.85 for the proposed method. The mean registration error of prostate centroids was 6.70 mm for the WSD method and 3.11 mm for our proposed method.

Conclusion: We developed a novel weakly-supervised, patch-based MR-US prostate registration approach. Through MRI-US registration, we could incorporate the MRI-defined target into real-time US imaging to guide HDR catheter placement. This can increase the chance of achieving the optimal tumor boost delivery, enable accurate boost planning and delivery, and hence ultimately improve prostate HDR treatment outcomes.

Keywords

Not Applicable / None Entered.

Taxonomy

TH- Brachytherapy: Registration

Contact Email