ABS 025: Towards the development of synthetic ultrasound data for image classification of Obstructed Defecation Syndrome

Nalima Munyofu ¹ , Ghazaleh Rostaminia ² , Steven Abramowitch ¹

¹ Department of Bioengineering, University of Pittsburgh, PA, USA
² Division of Urogynecology, NorthShore University HealthSystem, Chicago, IL, USA

Van Wickle (2025) Volume 1, ABS 025

Introduction: Obstructed defecation syndrome (ODS) without dyssynergia is a form of constipation associated with structural defects in rectal support. Diagnosing ODS remains challenging due to inter-patient anatomical variability. The overarching aim is to train deep learning models for automated diagnosis. However, clinical image acquisition and manual annotation are time-consuming and resource-intensive. This study investigates the feasibility of generating synthetic ultrasound images using a dual-stage pipeline. In the first stage, Houdini™ software was used to create geometric models representing the rectum, vagina, cul-de-sac, and levator plate using simple geometries. The 3D model was animated to mimic and simulate dynamics observed during ultrasound imaging. The second stage utilizes Field II, a MATLAB-based ultrasound simulation tool, to simulate collecting radio frequency data of anatomic phantoms based on the Houdini model. The radio frequency data was then processed to create synthetic ultrasound B-mode images. One simulation using 50,000 point scatterers in the phantom required approximately 20 minutes to collect 50 lines of radio frequency data. The results demonstrated the pipeline’s ability to produce realistic ultrasound images with clear anatomical boundaries. The collection of real clinical images is limited by budget and can span several months. The pipeline is more time efficient, and allows for scalability to create batches of models and images. Future work will improve the anatomic accuracy of the model, integrate patient-specific MRI-derived geometries, and refine parameters to match clinical probe specifications. This study establishes the pipeline as a viable method for synthetic ultrasound image generation.

Methods: This study leverages a dual-stage pipeline for generating synthetic ultrasound data with precise annotations. First, the 3D animation software Houdini™ (version 20.0, SideFX, Inc.) is employed to create and animate geometric models representing anatomical structures. These models serve as the ground truth for the synthetic data generation process. Subsequently, the radiofrequency simulation program Field II is utilized to transform 2D exports from Houdini™ into synthetic ultrasound images. Field II simulates ultrasound transducer fields and imaging using linear acoustics, capable of modeling emitted and pulse-echo fields for various transducers and simulating linear imaging of human tissue. The success criteria are for the pipeline to be more time efficient in comparison to acquiring real clinical images. The pipeline should allow for scalability to create large sets of data, and the images produced should be realistic.

Results: The model and thus ultrasound image is a simplified representation of ODS, however there are clear boundaries between the geometries in the image. Scatterers with higher amplitudes (i.e. higher density) produce stronger echoes when the simulated ultrasound waves interact, while those with lower amplitudes produce weaker echoes. Variation in amplitude contributes to the contrast and overall appearance of the ultrasound image, allowing different structures within the phantom to be distinguished. Despite the simplicity, the images capture the desired tissue behavior. The boundaries of the synthetic images match the movement and general shape of the segmentations on the clinical images.

Discussion: Field II is successful at simulating RF data and produced realistic ultrasound images based on the anatomical phantoms. The pipeline achieved all three of the success criteria: time efficiency, scalability, and relative realism. The pipeline poses a viable method to generate realistic ultrasound images. One limitation is the tradeoff between simulation accuracy and computation time. Increasing the number of scatterers and scan lines enhances image resolution but prolongs the process of RF data calculation. Future research will focus on developing the Houdini model by more accurately configuring geometry attributes. Future work may employ geometries obtained from MRI.