Authors (including presenting author) :
Chan JCY(1), Li AMK(1), Yu SJA(1), Chan DLH(1), Lui KWC(2), Wan JCH (2), Garifullin A(2), Lam YF(2), So BCH(2), Chung APM(2), Lee DPK(2), Chan JKY(2), Choi KH(1), Lam HY(1), Cheung CW(1)
Affiliation :
(1) Department of Radiology, Queen Mary Hospital, Hong Kong (2) Information Technology and Health Informatics Division, Hospital Authority Head Office, Hong Kong
Introduction :
Pulmonary embolism (PE) is a life-threatening condition requiring timely diagnosis via CT pulmonary angiogram (CTPA). The increasing demand for urgent imaging has lengthened turnaround times, delaying treatment. We propose developing a validated, explainable artificial intelligence (AI) system to triage and enhance radiology workflows, aiding radiologists in decision-making. The Pulmonary Embolism AI Initiative aims to create an in-house AI pipeline for interpreting CT images to improve patient care. The AI pipeline will screen CTPAs with a high negative predictive value (NPV) and detect right heart strain, flagging positive cases for immediate review and treatment. Additionally, it will provide explainable predictions for manual verification. This initiative also serves as a pilot program to explore collaboration between IT and Radiology teams for future advanced imaging AI developments.
Objectives :
1. Develop a family of deep learning AI models for PE detection on CTPA to streamline workflows and enhance patient outcomes. Key functionalities include: - Detection, localization, and characterization of PE - Identification of right heart strain - Explainability of predictions for radiologist verification 2. Upgrade Radiology AI Project Annotation Platform for axial imaging 3. Capacity building for advanced imaging AI development
Methodology :
Model development and validation were performed in HA on-premises GPU servers. A supervised deep learning model was chosen as the base, utilizing a public CTPA dataset from RSNA, labeled by expert radiologists, consisting of 9,446 studies (2.3million images). Transfer learning will tailor the model to local data. We retrieved 573 CTPA cases from QMH for training and validation, with 173 positive for PE. The Radiology AI Project Annotation Platform was enhanced for efficient CT slide navigation and labelling. Radiologists labeled a subset of cases to test the initial model's performance. Image labelling criteria and strategy were established to ensure accuracy and quality of labelled data.
Result & Outcome :
The base model achieved an AUC of 0.952 on the RSNA validation dataset, with a representative validation set achieving 0.920. The YOLOv11 model on lesion localization demonstrated a precision of 0.52 and recall of 0.47. For local dataset testing, the model accurately localized 69.2% of cases, identifying PE precisely while minimizing false positives, avoiding mislabelling of mimics of true PE, for instance. Further analysis has identified areas for model refinement and optimization.
