Data-efficient ultrasound imaging analysis with deep learning

Deep learning techniques for ultrasound imaging, which is a widespread non-invasive diagnostic imaging tool, have advanced recently. These techniques are designed to tackle long-standing challenges in ultrasound imaging, such as spatial resolution, motion analysis, and common image processing tasks (e.g., segmentation). The operator-dependent nature of ultrasound scans leads to inter-operator variability and analysis complexities. This makes the availability of large public ultrasound datasets and data annotations limited. The cornerstone of this thesis is the concept of data-efficient analysis. Data-efficient analysis aims to maximize model performance with limited data, thus addressing aforementioned challenges. The contributions of this thesis are manifold. Chapter 2 introduces the CCycleGAN model, which is a novel approach to generate ultrasound sector images with spatial resolution that is more spatially uniform throughout the entire sector field of view. As mentioned earlier, operator-dependency of ultrasound scanning makes acquisitions of paired high-resolution and low-resolution ultrasound images impractical. By leveraging the power of CycleGAN with unpaired data, CCycleGAN bridges the gap between unpaired ultrasound configurations (linear array vs. phased array), thus improving spatial resolution. The CCycleGAN model, with an efficient training strategy with a newly proposed constrained-consistency loss, is tailored to ensure that generated images retain critical anatomical details and speckle patterns. The optimized architecture also ensures swift inference time (an average of about 200 ms per image) suitable for near real-time clinical applications. Chapter 3 explores the field of 2D (i.e., lateral and axial) motion estimation in ultrasound imaging, with a special focus on the challenging lateral direction. The proposed TransPWCLite model, a lightweight transformer-encoded optical flow pyramidal network, is introduced for accurate 2D motion estimation. This model not only showcases the potential of transformer architectures in capturing temporal dynamics inherent in ultrasound sequences but also accelerate inference time (an average inference time of about 300 ms per image pair). The training with data augmentation ensures data-efficient learning, making it suitable for limited training samples. In Chapter 4, the proposed WMSEUNet model tackles tissue segmentation in dynamic lung ultrasound images. WMSEUNet is a weakly-supervised mask enhanced multi-scale self-attention efficient UNetIt integrates self-attention mechanisms to focus on critical regions in the ultrasound images, ensuring accurate segmentation. This is particularly beneficial when images exhibit intrinsic speckle noise, low contrast, and obscured tissue boundaries, such as those between cutaneous layers and muscles. The ability to train the model with weakly-supervised annotated data alleviates extensive annotations, making the training more feasible and efficient. Chapter 5 first summarizes the pivotal role of data-efficient deep learning models in ultrasound imaging, not only to address the intrinsic challenges associated with ultrasound imaging but also to provide useful insights into future research, with aspirations of widespread clinical integration. Lastly, the thesis highlights potential areas, including semi-supervised learning, model optimization, and integration with other modalities. In conclusion, this thesis emphasizes the importance and demonstrates the feasibility of achieving high deep learning model performance while relaxing data requirements in ultrasound imaging and analysis tasks, including spatial resolution improvement, motion estimation, and tissue segmentation.