Ultrasound is one of the most common medical imaging tools in the world because it’s safe, affordable, and shows results instantly. However conventional ultrasound mainly produces gray-scale images that only show how strongly tissues reflect sound, not what they’re actually made of. This limits its usefulness for detecting certain diseases, including breast cancer. In my PhD project, we are developing novel ways to measure the speed-of-sound inside tissue using conventional handheld ultrasound probes and standard clinical machines. Because different tissues, healthy or diseased, carry sound at different speeds, this gives a new, quantitative picture of the body that could improve diagnosis. Speed-of-sound reconstruction with handheld transducers is an ill-posed inverse problem. Although analytical methods exist, they require handcrafted regularization and carefully tuned parameters, and the resulting solutions are often computationally expensive. Deep learning has emerged as a powerful tool for learning data distributions. Classical deep-learning approaches (e.g., for image classification or segmentation) perform well when large annotated datasets are available. However, for medical image reconstruction, large datasets are typically scarce, and black-box neural networks may suffer from generalization. Model-based deep learning provides a promising alternative by integrating known imaging physics and models into the learning process, enabling training from limited data and/or simulations while still maintaining generalizability despite potential domain gaps. Within model-based deep learning frameworks, neural networks often act as denoisers. Recently, diffusion models have emerged as highly effective denoisers, suggesting that they can replace purely CNN-based denoisers in such frameworks. In this project, we will develop diffusion-model-based denoisers for speed-of-sound imaging within a model-based deep learning framework. Our initial experiments show promising results, and access to Alvis resources will enable further development and large-scale experiments. We will compare our method with both black-box and other model-based deep learning approaches. Our research group has several prior publications that establish key foundations for speed-of-sound imaging. This project will build on that work by advancing deep-learned solutions for this emerging ultrasound imaging technique. In our previous proposal, we were unable to fully utilise the allocated resources because the preliminary results were not yet promising. This time, however, we have generated new datasets with encouraging preliminary results, and we will make full use of the resources provided.