AI-integrated treatment planning and real-time guidance for ultrasound-guided high intensity focused ultrasound (USgHIFU)

High Intensity Focused Ultrasound (HIFU) is a non-invasive, non-ionising therapeutic technology that uses acoustic energy to ablate tumours at depth in the body. Compared with traditional surgery, it has the potential to reduce recovery time, minimise side-effects, and improve patient safety. However, challenges in treatment planning and imaging guidance have limited its widespread adoption.
This thesis presents an AI-integrated, ultrasound-guided HIFU system to address these challenges, combining real-time multi-organ segmentation with collaborative robotics (co-robotics) for patient-specific planning and guidance. These advances are demonstrated in the context of benign symptomatic thyroid nodules; however, the core methods and results developed could be broadly applied to a variety of treatment planning and ultrasound imaging tasks.
Initial work with lung ultrasound phantoms established lightweight Deep Learning (DL) models for live B-mode segmentation (mDSC = 0.74, 33 FPS), showing promise for future diagnostic systems. These methods were then extended to enable a co-robotic 3D ultrasound system, which achieved state-of-the-art geometric reconstruction accuracy on a purpose-designed quality assurance phantom (DSC-3D = 0.94 +- 0.01, HD95 = 1.17 +- 0.12 mm), approaching the spatial resolution limits imposed by ultrasound imaging.
In thyroid ultrasound, the co-robotic 3D ultrasound system allowed live multi-organ reconstruction and thyroid volumetry which closely matched the CT ground-truth (37.7 +- 1.3 mL vs. 37.9 +- 1.8 mL). This outperformed the current clinical standard ultrasound technique, with implications for radioiodine dosimetry and future semi-automated HIFU planning.
Finally, to support clinical translation, a new UK ultrasound dataset was curated, enabled by a novel data extraction tool which extracted key diagnostic information from free-text radiology reports (micro-average sensitivity, specificity, and F1-score of 0.97, 0.96, and 0.94).
Overall, this work represents progress towards addressing key barriers to wider HIFU adoption, including long treatment times, complex manual planning, and inter-operator dependence. Addressing these will be key to HIFU’s role in improving patient outcomes in the future.