Research summary · public version
How can AI learn from scientific images without losing the rules of the physical world?
I study generative and uncertainty-aware machine learning for polarimetric medical imaging. The goal is models that are expressive, physically coherent, and explicit about what they do not know.
1. Polarimetric imaging
A polarimetric image records how light interacts with a sample, not only its colour and brightness. That extra structure can reveal information that ordinary images miss.
2. Generative models
I explore variational autoencoders, flow-based models and related methods to represent complex scientific images, and to create useful variations when labelled data are scarce.
3. Physics-informed learning
A generated image can look convincing and still be physically impossible. My work studies how mathematical and physical constraints can be built into the representation, the decoder and the evaluation.
4. Uncertainty in low-data medicine
When datasets are small, uncertainty is part of the result. I compare what models predict, where they become unstable, and whether generated data actually help a downstream task.
What is not shared here
Patient data, identifying information, unreleased experimental details, and claims that still need independent validation.