Sen, Snigdha; (2025) Non-invasive Tumour Microstructure Estimation with VERDICT-MRI and Deep Learning. Doctoral thesis (Ph.D), UCL (University College London).

Text SnigdhaSen_FinalThesis.pdf - Accepted Version Access restricted to UCL open access staff until 1 January 2026. Download (28MB)

Abstract

Diffusion-weighted MRI (dMRI) is a powerful imaging modality that provides unique insight into the micro-arrangement of biological tissues by tracking the motion of water molecules. While dMRI offers useful clinical information for the non-invasive diagnosis of urological cancers, it lacks the detailed tissue characterisation available from histology. By combining dMRI with biophysical models that reflect the underlying tissue microstructure, similar information to histology can be inferred. However, these methods require significant improvement for clinical adoption. This thesis explores how computational modelling for microstructure estimation can be enhanced through advances in deep learning. Specifically, I utilise the VERDICT (Vascular, Extracellular and Restricted Diffusion for Cytometry in Tumours)-MRI framework, which combines a tailored dMRI acquisition protocol with a multi-compartment biophysical model for cancer characterisation. I show that these techniques are able to characterise elusive false positive cases of prostate cancer on multiparametric MRI, potentially reducing unnecessary biopsies. I demonstrate that parameter estimation for such models can be improved using self-supervised deep learning methods, resulting in better diagnostic performance. Extending this work beyond the prostate, I develop a VERDICT model for renal tissue to aid in differentiating renal cell carcinoma subtypes. I also introduce novel feature selection methods to optimise imaging protocols, improving clinical feasibility. Finally, I generalise the self-supervised model fitting approach across other dMRI models through an open-source software framework designed for broader research use. In summary, this thesis works to enhance the diagnostic potential of dMRI by integrating computational models and deep learning. Although focused on prostate and renal cancer, the advances in model fitting and protocol design are applicable for the microestimation of other cancers and clinical conditions, such as multiple sclerosis.

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as CSVXMLJSON

Archive Staff Only

View Item