A University of Pittsburgh researcher is developing artificial intelligence (AI) software to transform computed tomographic pulmonary angiography (CTPA) into non-contrast CT (NCCT) scans. This software will then be used to train an AI algorithm to detect pulmonary embolisms (PE) on routinely acquired NCCT scans. This novel software and AI algorithm could substantially improve the detection of life-threatening PE, offering a more accessible, efficient, and potentially safer method of diagnosis.
Description
PE, a clot that can block a blood vessel in the lung, can be a fatal cardiovascular emergency. In the United States alone, PE are associated with 370,000 hospital visits. As many as 10% of patients with PE die within an hour and 30% within one month. Early diagnosis and treatment are key to preventing these deaths and rapid screening techniques are an unmet clinical need. This novel AI-based approach could meet that need and accelerate early detection of PE, improving patient outcomes through reduction of mortality and long-term healthcare needs.
Applications
- Pulmonary embolism
Advantages
At present, there exists no simple PE diagnosis method as the key symptoms (e.g., shortness of breath and/or chest pains) can be non-specific and may result in misdiagnosis. The gold standard diagnostic tool is CTPA using iodinated contrast agents. However, not all patients are suitable for this approach particularly those with impaired renal function or allergies. While NCCT, a cheaper, accessible alternative, would overcome the challenge of contrast agents, the sensitivity is very low. Motion artefacts and a low signal-to-noise ratio make it currently unsuitable for PE diagnosis or screening.
This novel approach is designed to overcome the sensitivity issues of NCCT using AI-based assessment. Improvements in the ability of NCCT to diagnose PE could provide several benefits including the use of routine chest CT scans (for other diseases) to assess for PE, better accessibility to CT scanning particularly in emergencies without delays to test for allergies or kidney function, and reduced costs compared to CTPA. In the long term, AI assessment of NCCT scans could facilitate a screening program for PE in high-risk patient populations.
Invention Readiness
Currently, in the early stages of development, a CycleGAN model was trained to transform the chest CTPA scans into NCCT using a dataset of CTPA and NCCT paired images. A convolutional neural network model will be trained and developed using CTPA images of over 7000 scans (30.4% positive for PE) from the publicly available RSNA-PE datasets to detect PE. Once optimized, external validation will be performed using an independent dataset of 200 subjects (50% positive for PE) with paired CTPA and NCCT scans. Further large-scale optimization would be required before clinical use.
IP Status
Patent pending
