Study about deep Bifurcation learning model for separation of ultrasound echo signals and tissue acoustic properties

We are developing deep learning models for extracting specular and speckle signal components in medical ultrasound echo signals and for quantitatively estimating acoustic properties such as reflection and scattering. The model reported in this paper is based on an autoencoder, which is similar to the single-task model we have separately developed, it is characterized by the fact that it is a bifurcation multitask model, which is oriented to efficiently separate signal components and acoustic properties with high accuracy. The model parameters can be significantly reduced, for instance, 118.5M to 95M. Moreover, the model is improved by integrating the well-known pre-trained ResNet-50 encoder and employing the SSIM (structural similarity index measure) for better training and evaluation. In addition, a trainable shared loss is introduced to reduce the degradation of learning performance due to imbalance in the magnitude of loss for each task. As a result, the improved model outperforms the original single-task models in terms of signal component separation and acoustic property estimation with an average improvement of 0.352 and 0.279 in the SSIM metric, respectively.Clinical Relevance- This study is relevant for clinical use as it is able to give insight on the improvement of ultrasound imaging using deep learning. With further research, echo signals and acoustic properties could be more effectively utilized in clinical diagnosis.