¹û¶³Ó°Ôº

The aim of the module is to introduce solutions to practical challenges that arise when working with clinical and medical imaging data in machine learning/AI settings that go beyond the capabilities of classic machine learning algorithms. For instance:

• How can we make the best use of the limited labelled medical imaging data (as compared to the abundant data for natural imaging data)?
• How can we homogenise data coming from different collection centres (e.g., hospitals) or different machine manufacturers to ensure generalisability and (broad) applicability of these models?
• How do we compute and quantify the uncertainty of the recommendations?
• How can we generate ‘interpretable’ results that can be cross-checked by clinicians?
• How can we address the risk of bias in AI applications in the healthcare setting?

Each of these questions will be addressed from the angle of statistical modelling and modern machine learning in the context of clinical application and translation. More specifically, this will include:

• Data simulation and generation (e.g., dealing with missing data, dataset balancing, generating synthetic data, augmentation, simulation)
• Data homogeneisation and generalisation (e.g., meta-analysis, COMBAT, translation, domain adaptation and federated learning)
• Uncertainty modelling (e.g., Multiple sampling, Ensemble, Deep Learning Model uncertainty)
• Longitudinal and trajectory modelling (e.g., Mixed effects models, Gaussian Processes, Recurrent Convolutional Neural Network, ÌýLong Short-Term Memory)
• Interpretability and evaluation (Attention maps, evaluation metrics, surrogate validation)

Ìý