This thesis examines the use of magnetic resonance imaging (MRI) techniques in the detection of breast cancer and the prediction of pathological complete response (pCR) to neoadjuvant chemotherapy (NACT).

This thesis compares the diagnostic performance of diffusion-weighted imaging (DWI) models in the breast using a systematic review and meta-analysis. Advanced diffusion models have been proposed that may improve the performance of standard DWI using the apparent diffusion coefficient (ADC) to discriminate between malignant and benign breast lesions. Pooling the results from 73 studies, comparable diagnostic accuracy is shown using the ADC and parameters from the intra-voxel incoherent motion (IVIM) and diffusion tensor imaging (DTI) models. This work highlights a lack of standardisation in DWI protocols and methodology. Conventional acquisition techniques used in DWI often suffer from image artefacts and low spatial resolution. A multi-shot DWI technique, multiplexed sensitivity encoding (MUSE), can improve the image quality of DWI. A MUSE protocol has been optimised through a series of phantom experiments and validated in 20 patients. Comparing MUSE to conventional DWI, statistically significant improvements are shown in distortion and blurring metrics and qualitative image quality metrics such as lesion conspicuity and diagnostic confidence, increasing the clinical utility of DWI.

This thesis investigates the use of dynamic contrast-enhanced MRI (DCE-MRI) in the detection of breast cancer and the prediction of pCR. Abbreviated MRI (ABB-MRI) protocols have gained increasing attention for the detection of breast cancer, acquiring a shortened version of a full diagnostic protocol (FDP-MRI) in a fraction of the time, reducing the cost of the examination. The diagnostic performance of abbreviated and full diagnostic protocols is systematically compared using a meta-analysis. Pooling 13 studies, equivalent diagnostic accuracy is shown for ABB-MRI in cohorts enriched with cancers, and lower but not significantly different diagnostic performance is shown in screening cohorts.

Higher order imaging features derived from pre-treatment DCE-MRI could be used to predict pCR and inform decisions regarding targeted treatment, avoiding unnecessary toxicity. Using data from 152 patients undergoing NACT, radiomics features are extracted from baseline DCE-MRI and machine learning models trained to predict pCR with moderate accuracy. The stability of feature selection using logistic regression classification is demonstrated and a comparison of models trained using features from different time points in the dynamic series demonstrates that a full dynamic series enables the most accurate prediction of pCR.