Multiple steps are involved in radiology, from image acquisition to analysis by a radiologist. Artificial intelligence (AI) is a computational process mimicking human behaviors, whereas machine learning (ML) is the algorithm that implements this task without explicit programming. In radiology, AI-based solutions have been proposed to address various challenges, from image acquisition to tasks that depend on human cognition. The radiology framework consists of several steps: data acquisition, image pre-processing, quality control and assurance, data management, detection and quantification, and quantitative image analysis. Currently, these steps are performed using conventional algorithms and human observers. We propose the integration of AI in radiology to enhance the performance of these steps.
In summary, in this thesis, we successfully developed and implemented AI-based algorithms to enhance various aspects of the radiology framework. The thesis aimed to optimize image acquisition, reduce radiation dose and acquisition time, enhance image quality, and perform image correction and denoising. This resulted in faster, safer, and more efficient imaging procedures, improving qualitative and quantitative imaging accuracy. In addition, integrating AI-powered tools for image quality checks, artifact detection, and disentanglement improved quality control and assurance procedures. These advances led to improved quantitative accuracy, enhanced image quality and diagnostic confidence, and reduced image artifacts, making medical images more reliable for clinical and research applications. Furthermore, this study developed privacy-preserving sharing algorithms to address data privacy concerns and explored federated learning for secure and efficient model development across the different centers. This approach allowed for collaboration between radiology departments and centers while maintaining data integrity and security.
Moreover, this thesis successfully developed automated detection and segmentation algorithms for accurate, efficient, and consistent analysis of regions of interest from medical images. These algorithms demonstrated high accuracy and reduced errors in extracting quantitative metrics, leading to more precise quantitative analysis. This research also focused on creating robust, repeatable, reproducible image biomarkers to standardize image biomarkers to improve the performance of radiomic and radiogenomics models. In addition, this study achieved its goal of developing AI-based diagnostic and prognostic models, integrating radiological data with clinical, patient-derived, and multi-modal information using radiomics and radiogenomics modeling. The resulting diagnostic and predictive models could help in improving the management of patients toward personalized medicine. Overall, this thesis successfully harnessed the power of AI to improve different radiology frameworks, including data acquisition, image pre-processing, quality control and assurance, data management, detection and quantification, and quantitative image analysis, which ultimately enhanced patient care and promoted personalized medicine. The achievements of this thesis have the potential to advance the different stages of radiology to develop DTs to contribute to better healthcare outcomes.
