Workshop Outcomes
1. What is computer vision and image processing?
As we engaged with various healthcare leads, it became clear that there was mixed understanding around terms such as computer vision and a general, often vague, interpretation of AI. The workshop helped to clarify these concepts, highlighting that AI can work with many different types of data, while computer vision specifically focuses on the interpretation and processing of images. These images can be static, such as a scan of a body part, or dynamic, such as visualising blood flow through a vein over time. By analysing, interpreting, and extracting meaningful information from images, computer vision creates opportunities to support decision‑making in healthcare—for example, improving diagnostic accuracy or enhancing clinical workflows.
2. Use cases discussed
The workshop had a heavy focus on Ophthalmology. An interesting term that emerged was “Oculomics”, which is an emerging field that uses high end imaging and AI to analyse the eye to diagnose and monitor diseases. The eye does not only give insights into visual health but shows promise to be window to observe neural degeneration. Incredible strength of Ophthalmology research and AI innovation across Northern Ireland creates a strong opportunity for advancements. Other use cases extended to digital pathology, community accessible diagnostics, cardiology images and gait analysis. Emerging low cost and community accessible imaging also presents as an alternative to higher cost diagnostics which have longer waiting lists.
3. Data quality > data quantity
The conversations included discussions on AI common pitfalls. Existing systematic research reviews have highlighted a lack of models ready for clinical use, dataset issues, flaws in methodology and bias in the data sets used to build trustworthy models. Allowing uncontrolled or misplaced optimization can lead to AI tools being not reliable or generalizable. Further work is needed to improve data, validate procedures, document methods, and encourage interdisciplinary work. The future of AI in health care will depend on addressing these.
4. Challenges
Reproducibility is a key area identified in image processing in AI. Many of the studies made it impossible for other researcher to replicate or provide any explainability to the decisions being made. These issues can become even more challenging when factors such as variance in the images from different diagnostic equipment. Simple variances such as the colour of the output can have an impact on outcomes. Pre processing of data becomes a bigger topic to standardise the inputs to create more consistency.
5. Challenges
Accessing Data to develop better AI models was a detailed discussion. Despite 5 million NI digital pathology images being available, there are legal, ethical, technical, and institutional barriers to getting access to these and other relevant imaging data sets. While some of the barrier are designed to protect patients and manage risk, there was a shared consensus that much more could be done to provide a step change in patient care while still protecting patients.
6. Challenges
“Going small” as the next big thing emerged as a recurring theme across multiple demonstrations. It is widely recognised that AI’s energy requirements have grown exponentially and are becoming increasingly unsustainable. This poses significant risks in terms of climate impact, rising energy costs, and the long-term business case for AI investment. Several discussions offered insightful perspectives on developing lightweight AI frameworks capable of delivering strong performance while using fewer computational and energy resources.
6. Challenges
“Going small” as the next big thing emerged as a recurring theme across multiple demonstrations. It is widely recognised that AI’s energy requirements have grown exponentially and are becoming increasingly unsustainable. This poses significant risks in terms of climate impact, rising energy costs, and the long-term business case for AI investment. Several discussions offered insightful perspectives on developing lightweight AI frameworks capable of delivering strong performance while using fewer computational and energy resources.
