Image Synthesis with Weak Supervision Model for Brachytherapy of Prostate Cancer
Y Pang1*, Y Huang2, X Chen3, B Lian4, T Royce5, P Yap6, J Lian7, (1) University of North Carolina at Chapel Hill, Carrboro, NC, (2) University of North Carolina at Chapel Hill, ,,(3) University of North Carolina at Chapel Hill, ,,(4) Chapel Hill High School, Chapel Hill, NC, (5) Unc Healthcare, ,,(6) University of North Carolina at Chapel Hill, ,,(7) University of North Carolina, Chapel Hill, NC
Presentations
TU-GH-BRB-6 (Tuesday, 7/12/2022) 1:45 PM - 3:45 PM [Eastern Time (GMT-4)]
Ballroom B
Purpose: Prostate magnetic resonance imaging (MRI) offers excellent details of structures and tumors, but is normally unavailable in the operation room for the brachytherapy of prostate cancer. Trans-rectal ultrasound (TRUS) imaging is widely available but suffers from low soft-tissue contrast and tumor visualization capability. The prostate is delineated on the ultrasound image for treatment planning but often not on the diagnosis MR image. We propose a weakly supervised deep learning model based on unpaired data to generate a TRUS-styled image from a prostate MR image. This makes pre-treatment planning possible and removes the need for pre-operation ultrasound image acquisition.
Methods: We take a weakly supervised learning approach with content disentanglement using unpaired data. The disentanglement of image appearance and structure fidelity mitigates the complexity of the synthesis process and reduces the scale of the network. We further propose a structural constraint and an adversarial mechanism to handle non-linear projections of anatomical structures and dominant intraprostatic lesion (DIL) between two image modalities. The training uses forty ultrasound images from previous brachytherapy and sixty-one unpaired MR images. The testing includes six paired samples. The peak signal-to-noise ratio (PSNR), structural similarity metric (SSIM) and dice similarity coefficients (DSC) are used to compare the performance of our model with other state-of-the-art methods. Brachytherapy planning was repeated on five previously treated patients using synthetic CT.
Results: Our model generated the ultrasound-styled image with the content of the MR image. The SSIM, PSNR and DSC are 0.90 ± 0.02, 17.27 ± 1.71 and 0.88 ± 0.12, respectively, which are superior to other model performances. Brachytherapy on synthesis image shows comparable dosimetry to clinical plans.
Conclusion: We proposed a method that synthesizes TRUS images from unpaired MR images. The image quality measurement and planning study show it is suitable for prostate cancer brachytherapy treatment planning.
Funding Support, Disclosures, and Conflict of Interest: This project is in part supported by NIH 1R01CA206100.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email