Ultrasound is an important modality for medical intervention such as vascular access because it is safe, portable, and low-cost. However, ultrasound scanning requires trained sonographers who are scarce, and it can be challenging to perform ultrasound examinations in disaster or battlefield scenarios. This motivates us to automate ultrasound scanning. One significant challenge in automating ultrasound scanning is performing the scan on a highly curved surface while simultaneously maintaining proper contact with the surface to capture high-quality images. Another major challenge while automating ultrasound scanning is that if an ultrasound probe is pressed too hard against the skin it causes significant anatomical deformations. Subsequently, these deformations present a major challenge in the generalization of tasks like segmentation and registration in ultrasound images. Therefore, in this work, I aim to improve the autonomy of a robotic ultrasound system and improve the generalizability of ultrasound imaging algorithms to work at different force values.

 In the first part of the thesis, I present an enhanced strategy for ultrasound scanning using a robot with minimal expert guidance. Our methods demonstrate improved quality in the collected ultrasound images compared to existing approaches. In the subsequent section, I address the challenge of deformable registration in ultrasound images, particularly when these images are acquired at different force levels. Instead of relying on supervised learning methods, that require exhaustive and expensive ground truth calculations, I propose an unsupervised approach to estimate the displacement field between various stages of deforming anatomy in ultrasound images. Furthermore, by analyzing displacement field patterns for different compression forces, we develop an accurate physics model for ultrasound image compression. This model enables the generation of photo-realistic ultrasound images at different compression forces. Leveraging these photo-realistic images, I augment a vessel segmentation model to enhance its generalization capabilities, particularly for higher force values.