Challenges and Opportunities for AI in Medical Image Analysis


In our Zoom for Thought on February 15th, 2022, UCD Discovery director Prof. Patricia Maguire spoke with Kathleen Curran, an assistant professor in Diagnostic Imaging in UCD School of Medicine, on UCD Discovery's Zoom for Thought series, about, "Challenges and Opportunities for AI in Medical Image Analysis".  In case you missed it, here are our Top Takeaway Thoughts and a link to the video.



The Role


Kathleen leads the AI in Medical Image Analysis module in UCD’s MSc AI in Medicine. She studied radiography before completing her PhD in computer science and has “always been at the interface of medicine and computer science”. She supervises PhD students, oversees three commercialisation teams and directs the UCD machine learning in medical imaging research group ( This is “quite a diverse group”, including computer scientists, biomedical engineers and clinicians developing skills in machine learning. For her work in musculoskeletal injuries Kathleen works with sports institutes, clinicians, engineers and industry. She also works in cancer detection and rare lung disease detection - “quite a range of different machine learning applications in medicine”.




Interpretation of medical images is subjective and there can be “quite a lot of variation” between clinical interpretation even within one hospital.  

“What AI can offer is an opportunity to standardise some of the interpretation in terms of providing objective metrics, and actually quantifying some of the results.” 

Another challenge is ensuring that datasets used to build machine learning models are unbiased with regard to gender, diversity and ethnicity.

Kathleen is also interested in the “big area” of  “interpretability or explainability”. She and her research groups are developing methods to provide that explainability or interpretability to clinicians, “so that they can trust the models and the outcome of the models that we produce”. 

Another “huge challenge” is getting these machine learning models FDA, CE or MDR approved. 


Interdisciplinary Effort


Interdisciplinarity is “absolutely key” to Kathleen’s work. She gives the example of a research collaboration with Prof. Cormac McCarthy and the respiratory department in St Vincent’s University Hospital, Dublin. 

“We've just published an article on the use of machine learning techniques to predict risk of pulmonary embolism in patients. We have also received funding to look at developing an AI tool for rare lung disease screening.” 

Expert clinicians will help Kathleen’s team with “annotating or drawing regions of interest around key image features”. Then computer scientists and engineers will “develop the models that can subsequently be embedded into radiology workflows”.


Injury Prediction


Kathleen’s team is also developing prediction models for risk of injury. This would be beneficial, for example, when assessing the training and resting recommendations for specific athletes.  “Can we combine imaging features and injury statistics to see if we can predict risk of recurrent injury?” 


Deep learning or machine learning?


When researching their pulmonary embolism paper, Kathleen’s team tried some advanced machine learning models, along with more traditional machine learning models. 

“We found that we were getting accurate, transparent results with the standard machine learning methods. It's very easy to think, ‘Let's throw everything at this problem’. But there is no need if there is a simple solution with our traditional methods.”


Data Protection


Kathleen co-chairs UCD’s data protection impact assessment committee. Ensuring a patient’s data and history are protected within machine learning models is “really important”. There are “lots of measures” that can be taken to de-identify patient data, which is particularly relevant when dealing with rare diseases in a small country like Ireland.  

“It's not just about removing the name; it’s much more complex than that. If you have a cohort of rare lung disease patients and you know the sex of the patient and their age, then it might be possible to re-identify that patient. We can add measures such as noting the age range rather than the actual age of the patient.”




As for whether algorithms will ever replace radiologists, currently we “absolutely need humans-in-the-loop”. Clinicians need to be involved in the process of feeding back information to the machine learning model, “so that we can interpret and trust the results”. As the field progresses further “we will have more augmented intelligence” - people and artificial intelligence working together.  


This article was brought to you by UCD Institute for Discovery - fuelling interdisciplinary collaborations. 



Kathleen Curran is an Assistant Professor in Diagnostic Imaging in UCD School of Medicine and an Affiliated Principal Investigator in the Centre for Biomedical Engineering. She is a funded PI in the SFI centre for research training in machine learning, directs the UCD machine learning in medical imaging research group ( and leads an MSc programme on AI in Medical Image Analysis.  She is the co-founder of the Irish Diffusion Imaging Society (2009) and is lead guest editor for the special issue AI Enhanced Diffusion MRI in Frontiers in Neurology (2022).

Kathleen has secured over 2 million euros in research innovation funding since 2019 and was the recipient of the InterTrade Ireland Project Exemplar Award for International Academic-Industry Partnership with Axial3D in 2019. She is currently the recipient of three competitive Enterprise Ireland Commercialisation Funding Awards in the AI in Medicine space. She Co-Chairs the UCD DPIA committee, represents Ireland on the management committee for the GEMSTONE (Genomics in Musculoskeletal) EU cost-action and is the European bioinformatics working group leader for this action and has led numerous international advisory and review panels for the European Commission since 2010.