Skip to main content
2024 BMES Annual Meeting Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.
Back

Biomedical Imaging and Instrumentation

Biomedical Imaging and Instrumentation - Poster Session E

Poster F5 - Towards the Development of Synthetic Ultrasound Data for Image Classification of Obstructed Defecation Syndrome

Saturday, October 26, 2024
10:00 AM – 11:00 AM EST
Location: Exhibit Hall E, F & G

Introduction:  Obstructed defecation syndrome (ODS) without dyssynergia is a form of constipation characterized by the inability to evacuate rectal contents that is not due to impaired peristalsis [1]. Our previous research has shown that in many female patients, this condition is associated with a structural defect compromising rectal support, necessitating surgical intervention via rectopexy [2]. Intravaginal ultrasound imaging of this condition reveals increased mobility of the posterior cul-de-sac (the junction between the vagina and rectum) towards the rectoanal junction during increased intra-abdominal pressure [2]. However, diagnosis is challenging due to significant inter-patient anatomical variability, particularly in those with structural defects. The overarching aim of this research is to develop and train deep learning models to guide ultrasonographers during image acquisition, assist in automated measurements and feature extraction, and aid in the diagnosis of ODS resulting from structural rectal support defects.
Training deep learning models for medical image classification requires a large dataset of annotated images, including both positive (defect present) and negative (defect absent) cases. Traditionally, this necessitates access to an extensive database of patient ultrasounds, with clinical experts manually annotating each image frame—a time-consuming and resource-intensive process. The present study aimed to investigate the viability of developing a pipeline for generating synthetic ultrasound images with corresponding annotations, suitable for training AI models. This approach could potentially address the challenges of data scarcity and annotation bottlenecks in medical AI development.

Materials and

Methods:  This study leveraged a dual-stage pipeline for generating synthetic ultrasound data. First, the 3D animation software Houdini™ (version 20.0, SideFX, Inc.) was employed to create and animate geometric models representing anatomical structures, serving as the ground truth for the generated synthetic data. Tubes represented the rectum and vagina, a triangular prism for the cul-de-sac, and a curved tube for the levator plate. Houdini's Vellum soft body solver was used to mimic ultrasound-observed dynamics. To simulate tissue heterogeneity, we applied stochastic volume densities to each geometry. Cross-sections of these models were rendered as bitmap images using Houdini's volume slice and camera nodes.
The second stage of the pipeline used the radiofrequency simulation program Field II [3-5], through MATLAB, to convert bitmaps into synthetic ultrasound images. The simulation modeled a transducer array with a 3.5 MHz center frequency, 100 MHz sampling rate, 192 elements (64 active), 0.44 mm element width, 5 mm height, 0.05 mm kerf, and a 60 mm initial focus, incorporating Hanning window apodization and multiple focal zones from 30 to 200 mm. The output was a grayscale image balancing resolution, penetration, and practical constraints for this initial feasibility assessment.
At this stage of the research, we conducted a qualitative visual comparison between the synthetic images generated by our pipeline and clinical ultrasound images obtained from patients. This comparative analysis was used to identify areas for refinement and to guide the next iterations of our synthetic data generation process.

Results, Conclusions, and Discussions: Figure 1 presents the results of our data processing pipeline, featuring three B-mode ultrasound images generated using Field II. These images are derived from three frames of the Houdini simulation. While the model represents a simplified version of ODS, distinct boundaries between the simulated anatomical structures are clearly visible. The variation in scatterer amplitudes, corresponding to different tissue densities, results in a range of echo intensities in the simulated ultrasound interaction. Amplitude variation contributes to the image contrast and overall echotexture, enabling differentiation between various structures. For a single bitmap, representing a single ultrasound frame, approximately 20 minutes of computational time was required.
For comparative analysis, Figure 2 displays clinical ultrasound images acquired from a patient diagnosed with ODS without dyssynergia during abdominal muscle contraction. These images were obtained using a 3D X14L4 Endocavity Transducer operating at 12 MHz.

The developed pipeline demonstrates potential as a viable method for generating realistic synthetic ultrasound images. Our future research will aim to refine the Houdini model by incorporating patient-specific geometries derived from MRI data and experiment with tissue density and stiffness parameters to better simulate the physical properties of soft tissues. Next, we will work on optimizing Field II parameters to more closely align with the specifications of the ultrasound probes used in clinical settings. Enhancing image resolution through increased scatterer density and scan line count extends the RF data calculation time, making the trade-off between simulation accuracy and computational efficiency a critical consideration at this iterative stage of pipeline development.

Acknowledgements (Optional):  Funding was provided by the Swanson School of Engineering and the Office of the Provost at the University of Pittsburgh.