Eliminating Silent Failures in Precision Medicine: An Uncertainty-Aware Imaging Framework for Equitable Patient Stratification in Neurovascular Disease Validated Using Multi-omic Data and a National Imaging Databank

Background

Cerebral small vessel disease (SVD) is a major driver of cognitive decline and stroke underlying senescence and a myriad of neurodegenerative diseases, yet targeting the right therapeutic strategy to the right patient to slow the progress of the disease remains a critical challenge. While AI-powered models offer automated scoring of white matter hyperintensities (WMH) (i.e., its main neuroradiological signature) [1] and allow delineation of most of its markers [2,3], they frequently experience "silent failures" Theygenerate highly confident but incorrect assessments in the presence of different sources of biases: heterogeneity in imaging acquisition systems and protocols, and different levels of prevalence of diseases. This creates profound healthcare inequalities, as algorithms disproportionately fail on older hardware typically found in underfunded or rural clinics, and in the presence of rare diseases or complex clinical scenarios. 

Concurrently, clinical trials and personalized treatments are hampered by biological heterogeneity because traditional macrostructural imaging fails to capture indicatives of the underlying molecular endotypes driving disease progression. Dr. Lewandowski and his team identified that activated perivascular fibroblasts (PVF) secrete SPP1 (Osteopontin) and COL6A1 (Collagen Type VI Alpha 1), which physically accumulate in enlarged perivascular spaces (PVS), an SVD marker [4]. Because early PVS enlargement and initial WMH boundary margins are notoriously faint and ambiguous on MRIs, integrating uncertainty quantification (UQ) ensures these highly subtle, initial structural alterations are precisely quantified rather than missed as noise or imaging artifacts, opening a precise therapeutic window at the right time, and helping with patient stratification. We hypothesise that knowledge of patient-specific vascular endotypes can be encoded into our UQ algorithms as an inductive bias in a way that improves SVD marker quantification at the individual level. This multimodal integration of omics and neuroimaging for patient specific SVD burden quantification has, to the best of our knowledge, never been attempted before. Expanding the UQ imaging framework previously developed in our group and integrating it with evidence from vascular multi-omics can eliminate algorithmic blind spots and map macrostructural damage directly to specific biological pathways. 

Aims

This project aims to 1) Develop an uncertainty-aware AI-powered pipeline that flags silent failures and optimizes automated WMH, PVS and SVD scoring across diverse MRI scanner protocols, and 2) Validate the framework in research datasets from which brain images and circulatory proteomics are available, for example, the Lothian Birth Cohort 1936. By pairing UQ-derived imaging metrics with perivascular fibrobalasts-driven transcriptomic endotypes, this framework will bridge the gap between raw data and equitable clinical action. 

Training outcomes 

The student will move beyond basic machine learning to master the next generation of reliable, clinical-grade AI systems. 1) Gain expertise in advanced computational medicine & trustworthy AI, including UQ methodologies (ensemble methods and conformal inference to measure model confidence and establish rigorous, mathematically guaranteed safety boundaries), robustness and bias mitigation (domain adaptation techniques to evaluate and fix algorithmic bias under different settings and clinical presentations), and brain image processing (methods commonly used by the first supervisor). 2) Gain end-to-end expertise in handling complex, high-dimensional biological data, its interpretation, and use it to underpin the AI computational image processing pipeline. 3) Development of rare, highly valuable data science skills in multimodal data fusion required to build joint proteo-imaging networks, mathematically linking structural neuroimaging voxels to molecular profiles. 4) Experience in transforming abstract mathematical metrics into actionable clinical trial endpoints, learning how to use UQ-validated biomarkers to reduce required sample sizes and screen for therapeutic responders. 5)  A deep understanding of how to audit technologies for systemic healthcare inequalities. The student will learn how to design medical algorithms that actively protect vulnerable patient populations diagnosed on resource-constrained clinical infrastructure. 

References

1. doi: 10.1016/j.media.2025.103697 
2. doi: 10.1007/978-3-031-48593-0_1 
3. doi: 10.1016/j.cmpb.2025.108715 
4. doi: 10.1038/s41591-021-01295-9