Coronary artery calcification (CAC) scoring is a reliable tool used to assess the likelihood and severity of cardiovascular disease. Radiologists must evaluate the position, density, and volume of calcium deposits in a heart by reviewing CT scans and assigning a total calcium score. These CAC scores are an important factor in clinical decision-making.
The calcified areas that radiologists aim to identify are very small, and even from an expert perspective they can be easy to miss. Minimizing the oversight of any calcium deposits is a critical aim that directly helps clinicians and impacts patients.
The HeartLens project is an EXCEL Research Committee effort bringing UK Divisions of Radiology & Cardiology, the Computer Science Department, and CAAI in collaboration to develop an AI-aided diagnostic tool. The tool exists to help radiologists review CT scans, assess coronary artery calcification, and make recommendations to cardiologists that inform diagnosis and treatment plans. The development of such a tool enables earlier detection of calcium deposits which means that interventions are more timely and treatment options can be less invasive/aggressive.
Using the DINO Transformer Model with CT scan datasets from Stanford (a public dataset) and the UK, feature extraction identifies calcified areas within arteries, and visualizing these features highlights regions for closer clinical review. Color coding indicates the severity and importance of each calcified area. Associated colors signify severity, and therefore importance. Locating, highlighting, and labeling these features gives clinicians more information, empowering them to detect risks earlier on. Additionally, these extracted features can support tasks such as classification, segmentation, and integration with large language models (LLMs) to further enhance diagnostic accuracy and utility.
Literature review indicates this research is on the cutting-edge and the group are pioneers in this area of work. The long-term scope of this project is not limited to CAC scoring and cardiovascular disease alone; other heart diseases and conditions could be extracted and highlighted with this model. The tool in development has broad applications for better early detection and diagnosis using CT scans. In the future, the tool developed through this research will be packaged in an easily accessible web-format that can integrate the data from CT scans with a variety of machine learning tools available withing CAAI’s CLASSify, further promoting research and supporting clinical decision making.
