Tools
Tools
Loukil, Zainab 
ORCID: 0000-0003-2731-7051
(2024)
A Hybrid Approach to Intelligent
Prediction of Medical Conditions A Framework for Advancing Medical Diagnostics through Novel Hybrid Deep Learning Models
DenCeption and HyBoost for Enhanced Feature Extraction and Predictive Accuracy in Medical
Image Analysis.
PhD thesis, University of Gloucestershire.
doi:10.46289/8XU8FE24
|
Text (Final Copy)
14692 Loukil (2024) PhD_A Hybrid Approach to Intelligent Prediction of Medical Conditions.pdf - Accepted Version Available under License All Rights Reserved. Download (18MB) | Preview |
Abstract
Medical image analysis is currently challenged by the need to achieve precision in diagnostic methods while maintaining broad applicability across diverse datasets. This challenge is intensified by the intricate details present in high-dimensional medical imaging data, affecting diagnostic effectiveness and patient care. Precise feature extraction is crucial in identifying patterns vital for medical diagnoses, yet current models often struggle with real-world variability, including diverse imaging conditions and patient demographics. This research advances the field by introducing a novel hybrid feature extraction framework, DenCeption, and a predictive model, HyBoost, which address these challenges through improved disease prediction accuracy, generalisability and adaptability. DenCeption, combining DenseNet-169 and Inception-V4 architectures, achieved a notable accuracy of 91.3%, surpassing existing models like DenseNet-121 (89.3%) on the BRATS MRI dataset, demonstrating its superior feature extraction capabilities. The hybrid feature extraction framework also proved adaptable across multiple datasets, including MRI and Retinal images, with accuracy reaching 98.9% in the Retinal dataset, significantly outperforming traditional methods. HyBoost, integrating multiple machine learning algorithms and leveraging patient demographic and physiological data, further enhances predictive accuracy. For instance, on the OCT dataset, HyBoost achieved an accuracy of 98.33%, with a sensitivity of 99.45%, outperforming existing models like XGBoost and AdaBoost. These improvements are supported by extensive testing across various datasets, Fundus, OCT, and X-ray, where HyBoost consistently demonstrated superior performance metrics, including low mean absolute error and high precision. Moreover, a new evaluation mechanism, involving a sophisticated performance measurement matrix (PMM), systematically selects the most optimal evaluation metrics, ensuring the robustness and clinical applicability of the models. This mechanism addresses the limitations of existing evaluation approaches, further enhancing the interpretability and reliability of the developed models. This research represents a significant advancement in medical image processing, setting new benchmarks in predictive medical imaging analytics. By systematically improving model performance and integrating advanced machine learning and deep learning applications, this work revolutionises medical diagnostics, achieving high accuracy rates and robust disease prediction across multiple imaging modalities.
| Item Type: | Thesis (PhD) | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Thesis Advisors: |
|
|||||||||
| Uncontrolled Keywords: | Medical image analysis; diagnostic methods; broad applicability; diverse datasets; diagnostic effectiveness; patient care; hybrid feature extraction framework; DenCeption, predictive model; HyBoost, | |||||||||
| Subjects: | Q Science > QA Mathematics > QA76 Computer software R Medicine > R Medicine (General) R Medicine > RA Public aspects of medicine > RA407 Health status indicators. |
|||||||||
| Divisions: | Schools and Research Institutes > School of Business, Computing and Social Sciences | |||||||||
| Depositing User: | Anne Pengelly | |||||||||
| Date Deposited: | 15 Jan 2025 11:11 | |||||||||
| Last Modified: | 16 Jan 2025 16:15 | |||||||||
| URI: | https://eprints.glos.ac.uk/id/eprint/14692 |
University Staff: Request a correction | Repository Editors: Update this record
