Dr Nicola Dinsdale gave us an overview of the technical and logistical challenges facing the clinical translation of big data imaging studies.

Discussing research into the estimation of patient’s brain’s age and their true age using the presence of bio markers and white matter hyper-intensities, Dr Dinsdale explained the requirement for images to be labelled by experts has limited the availability quality data for training. We also learned about how the method of data acquisition has a major effect on the variability seen in medical images, with the same scanner being able to produce images with different variations on the same patient on the same day. Dr Dinsdale referred to this as the harmonisation problem which has led to research in to models that can account for/remove the variations cause by the image’s acquisition while retaining the biological variation in the image. She then moved on to discuss the inherent identifiability of medical images and the problems this entails when using them for training. Federated learning was presented as a solution, enabling models to train on local private data before aggregated training weights are taken to provide a generalised weighted model. Model interpretations was then raised as an issue, as clinicians cannot entirely trust a Blackbox’s decision making. The example of a gaussian distribution being presented to a brain age prediction model, resulting in the output of the average age of the training data was shown to emphasize the need for interpretability. Finally, an all-too-common issue in medical AI was discussed when it came to model evaluation, as we require high quality Gold Standard data to properly evaluate our models, but the limited available labelled data hinders accurate evaluation.

After Nicola’s talk we went into our Rate My Plot - The Musical session hosted by Bailey Andrews. In this session attendees gave brutal criticism on plots whilst enjoying a playlist composed by AI. You can enjoy the playlist here and see some of the beautifully critiqued figures below:

Blog written by: Alexander Coles