Exploring the mathematical relationship between medical images and clinical labels can elevate clinical decision-making.  


Hi! I am ESR4, Neel Kanwal, and I will describe a bit of what makes it harder to be entirely dependant on AI.


Automation in healthcare relies on the digitization of clinical data. Large datasets are vital in creating AI models for clinical assistance. Accumulating sufficient medical images pertaining to various diseases is not a straightforward task. The usecases and adaptation of machine learning methods have seen a steep rise in medical image classification and segmentation. It was all possible due to high-end computational resources, availability of data, and optimization methods. However, this frequentist approach may easily overfit or risk introducing bias and hinder achieving higher prediction performance. It makes clinicians hesitant to rely on model predictions entirely.

Modeling uncertainty of predictions can provide a competitive edge against conventional machine learning methods and makes the automated system more accountable for mission-critical tasks. Bayesian methods can facilitate establishing casual models by incorporating domain knowledge as a prior distribution. They can be interpreted as probabilistic mappings from input images to their output clinical label. The Gaussian process has analytical properties for approximating the entire functional space of input images and marginalizing it to predict new images. Gaussian process quantifies uncertainty by modeling the error of labels. Predicted outputs along with confidence intervals can ascertain better decisions in healthcare environments. This approach also works well for smaller datasets that are common due to certain rare medical conditions. I will work on that during my ongoing secondment at the University of Granada, Spain. I will update you on my technical and cultural learning outcomes from this secondment. Stay well till next time!

First visit to the city of Granada

Neel Kanwal – ESR4