Recognition, retrieval, and harmonisation for multicentre clinical data analysis
File(s)
Author(s)
Nan, Yang
Type
Thesis
Abstract
Artificial Intelligence (AI) has emerged as a transformative force in medical imaging, significantly enhancing how professionals interpret complex data. Persistent issues mainly include (1) the inherent complexity and clinical variability of medical images, demanding highly adaptive AI for precise recognition and interpretation. (2) Developing effective AI solutions requires large, well-annotated datasets, a process that is time-consuming and resource-intensive for clinicians. (3) Research has largely concentrated on the preliminary stages such as diagnosis and feature extraction, neglecting the AI's potential for prognosis or long-term disease outcomes. (4) Variations in imaging devices, patient status, and protocol specifics limit the reproducibility and generalizability of AI models. This PhD program addresses these obstacles by concentrating on computational methods for clinical tasks such as recognition, retrieval, and data harmonization across multicenter datasets. Initial projects employed supervised learning to assess model performance in complex scenarios, including fine-grained recognition in kidney pathology and airway segmentation in lung CT scans. Innovations introduced include a fine-grained recognition model with uncertainty evaluation for differentiating different glomeruli and a fuzzy attention neural network enhanced by adversarial learning for more accurate airway segmentation, outperforming standard 3D segmentation approaches. To address the scarcity of annotated data, the research explored unsupervised learning with a novel deep mixture model that incorporates constraints to prevent segmentation errors. The AIIB23 challenge, held in conjunction with the MICCAI 2023 conference, established benchmarks for AI in airway segmentation and mortality prediction, helping to identify current research deficiencies. Furthermore, a clinical study involving patients with lung fibrosis compared the efficacy of image-derived biomarkers with conventional clinical metrics in mortality prediction, highlighting AI’s potential in clinical applications. The program also explored computational data harmonization, introducing an image retrieval system to categorize and align images based on textural similarities, promising advances for future large-scale clinical studies.
Version
Open Access
Date Issued
2024-02-07
Date Awarded
01/09/2024
License URL
Advisor
Yang, Guang
Walsh, Simon LF
Sponsor
The CHAIMELEON Project
Grant Number
H2020 (952172)
Publisher Department
Bioengineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)