Yang, Qianye; (2024) Image Registration and Morphological Analysis in Prostate Cancer Imaging. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Yang_ID_thesis.pdf - Accepted Version Download (10MB) | Preview

Abstract

Prostate cancer is one of the most diagnosed cancers in males in many parts of the Western world. Early cancer detection is crucial for patients placed in an active surveillance project. My purpose aims to develop automatic medical image regis tration and morphological prediction algorithms, which could potentially facilitate MRI-based prostate cancer diagnosis. In the first project, it introduced an algorithm for longitudinal image registration, aiming at quantifying changes in regions of interest between a pair of images from the same patient taken at distinct intervals. By using both intensity-based similarity and gland segmentation for weak supervision, my approach markedly reduces the errors in target registration (TREs) on new patient images and outperforms the traditional iterative registration method. Additionally, the study conducts a detailed evaluation of various methods for sampling longitudinal data and introduces an innovative regularisation approach that leverages the maximum mean discrepancy to regularise training image pairs that have been sampled differently. A data set of 216 3D MR images from 86 patients were used in this study. The result of a mean TRE of 5.6 mm demonstrates statistically significant differences between the different training data sampling strategies. In the second project, it described a new framework for forecasting prostate morphological changes. An effective MR image registration method that uses feature-based alignment to match the capsules of the prostate gland, allowing for the measurement of shape changes through the application of dense displacement fields (DDFs). In addition, the method of kernel density estimation (KDE) is used to predict the future morphological changes represented by DDFs, bridging the gap between current and forthcoming data points before the actual future data is accessible. This KDE approach is innovative in its use of a distance metric that incorporates considerations of morphology, stage-of-progression, and duration-of change, which are critical for patient-specific predictions. The proposed method was tested on image masks that the registration network had not previously seen, avoiding the need for data from future time points. The validation was carried out on a longitudinal dataset with 331 images from 73 patients, resulting in an average Dice coefficient of 0.865 for the comparison between the ground truth and the prediction from the proposed algorithm. In the third project, it proposed a cross-modality registration approach, which utilises extra modality available only during training. Specifically, I aim to align intra-subject multiparametric MR images, focusing on the registration between T2-weighted scans and diffusion-weighted scans with high b-value. The proposed method utilizes the zero b-value scans in training to improve the registration between T2-weighted and high-b diffusion MRIs. Experiments were conducted on a dataset of 369 multiparametric MRI sets from 356 prostate cancer patients demonstrate a significant reduction in median target registration error from 7.96 mm to 4.34 mm. The findings highlight the effectiveness of the proposed learning-based networks in achieving accurate registration, outperforming classical iterative and other deep learning based methods that do not use the additional modality.

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as JSONCSVXML

Archive Staff Only

View Item