Deep learning-based automatic prostate segmentation in 3D transrectal ultrasound images from multiple acquisition geometries and systems

Nathan Orlando; Derek J. Gillies; Igor Gyacskov; Aaron Fenster

doi:10.1117/12.2549804

Translator Disclaimer

16 March 2020 Deep learning-based automatic prostate segmentation in 3D transrectal ultrasound images from multiple acquisition geometries and systems

Nathan Orlando, Derek J. Gillies, Igor Gyacskov, Aaron Fenster

Author Affiliations +

Proceedings Volume 11315, Medical Imaging 2020: Image-Guided Procedures, Robotic Interventions, and Modeling; 113152I (2020) https://doi.org/10.1117/12.2549804
Event: SPIE Medical Imaging, 2020, Houston, Texas, United States

Abstract

Transrectal ultrasound (TRUS) fusion-guided biopsy and brachytherapy (BT) offer promising diagnostic and therapeutic improvements to conventional practice for prostate cancer. One key component of these procedures is accurate segmentation of the prostate in three-dimensional (3D) TRUS images to define margins used for accurate targeting and guidance techniques. However, manual prostate segmentation is a time-consuming and difficult process that must be completed by the physician intraoperatively, often while the patient is under sedation (biopsy) or anesthetic (BT). Providing physicians with a quick and accurate prostate segmentation immediately after acquiring a 3D TRUS image could benefit multiple minimally invasive prostate interventional procedures and greatly reduce procedure time. Our solution to this limitation is the development of a convolutional neural network to segment the prostate in 3D TRUS images using multiple commercial ultrasound systems. Training of a modified U-Net was performed on 84 end-fire and 122 side-fire 3D TRUS images acquired during clinical biopsy and BT procedures. Our approach for 3D segmentation involved prediction on 2D radial slices, which were reconstructed into a 3D geometry. Manual contours provided the annotations needed for the training, validation, and testing datasets, with the testing dataset consisting of 20 unseen 3D side-fire images. Pixel map comparisons (Dice similarity coefficient (DSC), recall, and precision) and volume percent difference (VPD) were computed to assess error in the segmentation algorithm. Our algorithm performed with a 93.5% median DSC and 5.89% median VPD with a <0.7 s computation time, offering the possibility for reduced treatment time during prostate interventional procedures.

Citation Download Citation

Nathan Orlando, Derek J. Gillies, Igor Gyacskov, and Aaron Fenster "Deep learning-based automatic prostate segmentation in 3D transrectal ultrasound images from multiple acquisition geometries and systems", Proc. SPIE 11315, Medical Imaging 2020: Image-Guided Procedures, Robotic Interventions, and Modeling, 113152I (16 March 2020); https://doi.org/10.1117/12.2549804

ACCESS THE FULL ARTICLE