Towards Clinical Adoption of Computer Aided Detection of Pulmonary Embolism on CT Pulmonary Angiography

In recent years, computer-aided detection (CAD) of pulmonary embolism (PE) on CT pulmonary angiography (CTPA) has undergone significant advancements, resulting in its increasing use in daily clinical practice. As radiologists' workload rises, CAD utilisation helps to reduce the risk of missing (mainly incidental) PE. Especially CAD applications powered by artificial intelligence (AI) have been increasingly studied to determine diagnostic performance and assess the risks and benefits of their use. This thesis investigated the potential of CAD to optimise PE detection and the barriers hampering implementation in clinical practice.

Part 1 - Optimisation of CAD for pulmonary embolism diagnosis
We aimed to optimise the diagnostic performance of an accessible, highly sensitive CAD tool for PE detection, which relies on automatic segmentation of vascular structures and intraluminal HU values. Its low specificity, however, limited clinical use, as it increased radiologists’ reading time. To determine whether image quality, expressed as contrast-to-noise ratio, could improve diagnostic accuracy, we compared CAD performance on images with different noise reduction levels using hybrid iterative reconstruction (chapter 2). We observed higher specificity with greater noise reduction, though with a slightly lower (not significant) sensitivity. Radiologists should therefore be aware of a potential increase in missed PE when using CAD on highly noise-reduced images. Since higher noise reduction improved specificity, we investigated whether novel noise reduction technology could further enhance CAD (chapter 3). Paired conventional and virtual monochromatic images at 60 keV from a spectral-detector CT were analysed, showing significantly higher CAD specificity on virtual monochromatic images while maintaining high sensitivity in both groups.

Part 2 - Validation of AI for pulmonary embolism diagnosis
During our research, major advances in CAD and AI technology emerged. To assess the additional value of AI in clinical practice, we retrospectively analysed CTPA data from a large patient cohort using a commercially available algorithm (chapter 4). Although both the AI and radiologists missed PE cases, AI sensitivity and specificity were significantly higher than the initial radiology report, suggesting that combining AI with radiologists yields the best diagnostic accuracy. We repeated this study for routine contrast-enhanced chest CTs in the portal venous phase using an algorithm trained to detect incidental PE (chapter 5). Radiologists missed a significant proportion of incidental PE detected by the algorithm, while AI had slightly lower specificity than in the CTPA study. This confirmed AI’s added value in clinical practice, though physicians should remain cautious about overtreating small, asymptomatic incidental PE.

Part 3 - Implementation of AI for pulmonary embolism diagnosis
Several barriers hinder AI adoption in PE detection, including limited physician knowledge and acceptance, costs of implementation, and sustainability amid rapid radiological advancements. To assess sustainability, we analysed the performance of a commercially available algorithm trained on conventional CT images applied to spectral-detector CT data (chapter 6). Diagnostic accuracy did not significantly differ between virtual monochromatic and conventional images, suggesting AI sustainability, though continuous performance monitoring remains crucial. Physicians’ limited confidence may stem from the lack of transparency in commercial algorithms and the time and cost needed for clinical implementation. Developing in-house or open-access AI models may overcome these barriers. In our hospital, we implemented an open-access model tested on multicentre CTPA data (chapter 7). Trained on a public database with openly available code, it ensured transparency regarding both the model and training data. The model performed well on our hospital data but showed lower accuracy on spectral-detector CT, likely due to training on an older, smaller dataset with CT images from 1998. In contrast, the commercial algorithm maintained performance, indicating that more extensive training is required to ensure robustness and adaptability to evolving technologies.