Uncertainty Quantification in Medical Image Analysis

The “black box” problem in medical imaging

In oncology and radiology, segmenting structures — drawing exact boundaries around tumours — is a key step in treatment planning. Deep learning automates this well, but usually as a black box: it returns a segmentation mask without communicating how confident it is. On fuzzy boundaries or poor-quality scans, a standard network can be wrong while appearing fully certain.

My research aims to build models that know what they don’t know.

Deep evidential networks and belief functions

To close the gap between raw accuracy and clinical trust, my work integrates belief-function theory into deep networks. These deep evidential networks output more than a boundary: they produce a spatial uncertainty map. In lymphoma segmentation from 3D PET-CT, rather than forcing a decision on every pixel, the network assigns a “mass of belief”; where the PET and CT evidence conflicts or is ambiguous, it allocates belief to “ignorance” — in effect telling the clinician which regions to review.

Contextual discounting and multi-modality fusion

Diagnosis often fuses several imaging modalities (e.g. T1, T2, and FLAIR MRI), but not every scan is equally informative for every tissue type. I introduced fusion mechanisms such as contextual discounting, which let the model assess the reliability of a given imaging context and down-weight a modality corrupted by patient motion or artefacts before fusing it with cleaner data. Because expert pixel-level annotations are costly, I have also applied these belief-function methods to semi-supervised learning, achieving accurate brain-tumour segmentation with limited labelled data.

Foundations for trustworthy AI

Beyond individual architectures, I have helped structure the broader area of reliable AI: my 2024 review in Medical Image Analysis covers both probabilistic and non-probabilistic methods for uncertainty quantification and serves as a reference for researchers in the field. The aim across this work is to make automated diagnostic tools transparent, safe, and interpretable partners in patient care.