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.

Biomedical Imaging and Instrumentation

Biomedical Imaging and Instrumentation - Poster Session C

Poster G3 - Machine Learning for Estimating Acoustic Scatterer Size in B-mode Ultrasound Images

Friday, October 25, 2024
10:00 AM - 11:00 AM EST
Location: Exhibit Hall E, F & G

Introduction:  Current clinical standard B-scan ultrasound (US) imaging has been primarily used to visualize underlying tissue and assess health and function of various soft tissues. B-scan US images are generally reconstructed by harnessing the mechanical waves reflected from an acoustic source. A major challenge in B-scan US imaging is unfolding hidden patterns and visualization of acoustic scattering functions. Over the last few decades, several quantitative approaches have been integrated with US imaging that explore the potential of B-scan US to quantify underlying tissue properties, such as scattering size. While research has been promising, extracting this information from ultrasound images can be computationally demanding which limits real-time imaging applications. To address this need, we have developed an integrated approach of utilizing B-scan US images with a convolutional neural network (CNN) architecture to quantify average scattering size. The CNN network was modified based on the MobileNet architecture for the classification of ultrasound images and estimation of tissue characterization parameters.

Materials and

Methods:  A set of three homogenous tissue-mimicking phantoms were created using a gelatin and water solution. Each custom phantom was embedded with a fixed concentration of monodisperse ultrasound scatterers of size 15, 30, or 40 μm (diameter). Grayscale US images were then acquired from each phantom using a preclinical ultrasound system (Vevo F2, FUJIFILM VisualSonics Inc) equipped with the ultra high frequency 46 MHz linear array transducer. This 256 element transducer operates at a 32 MHz center frequency and was used for focused imaging. The final distribution of images collected was approximately 2500 each for the respective 15 μm, 30 μm, and 40 μm classes.
Using this data, a custom MobileNet transfer learning model was created with the goal of image classification based on the underlying ultrasound scatterer size. The final model utilized MobileNet as the base layer with four additional dense layers. Prior to training, images were cropped to the final size of 51x51 pixels (approximately 0.63 square millimeters). Note that cropping the ultrasound images serves the purpose of making the machine learning model more generalizable to small regions from non-homogeneous images. Data was split into training (60%), validation (20%), and test (20%) sections, and the CNN model was developed to maximize scatterer size classification accuracy. Techniques such as data augmentation, dropout, and L2 regularization were used to increase the robustness of the machine learning model and improve its accuracy during testing on the hold-out data.

Results, Conclusions, and Discussions: A custom MobileNet transfer learning model was developed and applied to ultrasound images acquired from a series of homogeneous phantoms embedded with different-sized acoustic scatterers. During testing on the hold-out data, the model produced accuracy and average area under the curve (AUC) values of 0.95 and 0.99 respectively. This exceeds the baseline accuracy for random sampling (33.3%), making a CNN a promising choice for the rough estimation of acoustic scatterer size. The overall precision of this model was approximately 0.95. Individually, the precision for 15, 30, and 40 μm classifications was 1.0, 0.92, and 0.93, respectively. The higher precision with which 15 μm images are classified is expected since there is a greater absolute difference in size between the 15 and 30 μm-sized scatterers compared to the 30 and 40 μm-sized scatterer pairs, making the 15 μm target the easiest of those studied for model classification. Notwithstanding, results from the custom transfer learning model applied to ultrasound images collected from homogeneous phantoms with 30 or 40 scatterers were encouraging. In practice, this model would be used as a sliding window on an image, predicting scatterer sizes for each pixel individually throughout a non-homogenous image. This would be used for applications such as scatterer-based segmentation and the quantitative assessment of healthy and diseased tissue. Future work will explore the impact of training image size and use of more complex phantom material types.

Acknowledgements (Optional):