Application of artificial intelligence in diagnostic medical imaging : from research to clinics

Abstract

Artificial intelligence (AI) and deep learning have shown promising results in numerous medical imaging tasks in recent years. However, challenges remain in the development and integration of AI systems in clinical practice. First, there is a lack of large and high quality, well-annotated medical imaging datasets for the training of deep learning models. Second, the evaluation of the medical image-based deep learning model often focuses on model performance and diagnostic accuracy. The validation or evaluation of the clinical performance of the models for their intended applications is lacking. Third, more consideration is needed for integrating AI models into the existing radiologist system and workflow. The Integration of AI models in terms of the processing of DICOM images and presenting prediction results is essential to ensure clinical utility without creating an extra burden for clinicians. This thesis explores and investigates how to address each of these challenges.

To address the lack of high quality labelled medical datasets, a labelled scheme was proposed that utilises DICOM metadata to extract labels relevant to the appearance of the images, such as imaging modalities, anatomic view, contrast agent, slice spacing, and body coverage. The study demonstrated that the weights of a 3D convolution neural network trained with the label extracted when transferred onto a liver segmentation task achieved higher performance compared with training from scratch. In addition to transfer learning, the use of hand-crafted quantitative radiomic features without the need for deep learning was explored. A support vector machine model incorporating tumour stage and pre-treatment MRI radiomic features was developed to detect patients with nasopharyngeal carcinoma at risk of 3-years disease recurrence after initial radiotherapy treatment. The model achieved high discriminability of patients with 3-years disease-free survival (AUC = 0.80) in both the training and independent test sets. However, it was found that many radiomic features were sensitive to variation in interobserver and image processing, which limits the model’s clinical utility. To evaluate the clinical impact of AI models in clinical application, a multi-modal AI strategy incorporating deep learning analysis of chest x-rays and emergency physicians’ chest x-ray findings reports was developed for selecting cases for the second screening of lung mass and nodules in the emergency department. A retrospective simulation study on multiple hospitals was conducted to evaluate the clinical performance of the strategy. The strategy was shown to reduce the number of cases to be screened by half while still managing to capture most potentially missed cases compared with only using imaging models. For the integration of the image-based model, a framework was proposed with a single unified preprocessing pipeline that characterises all incoming images intended for AI inference to provide information for series selection to downstream AI models used for clinical purposes.

In conclusion, this thesis explores different aspects of the development of AI applications in medical imaging. The results and contribution of each study are expected to facilitate the development and integration of AI systems into clinical practice.