The Optellum artificial intelligence prediction model has been demonstrated to outperform the best currently used model.
Prediction of the likelihood of malignancy is central to the clinical management of suspicious pulmonary nodules, which are lesions detected on CT scans that may be benign or the earliest stage of lung cancer.
In current practice pulmonologists typically rely on their judgement and experience to assess the risk of malignancy. A number of clinical risk models have been proposed to improve and standardise predictions, typically taking into account parameters measured from the scan (such as nodule size) and patient history. One of the mostly widely accepted clinical risk models is the Brock model, also known as the PanCan model, recommended by the British Thoracic Society guidelines for pulmonary nodule management.
A paper published in Thorax , explains the significance of the Optellum Lung Cancer Prediction success at risk prediction and its potential applications. Lead author, Professor David Baldwin is Chair of NHS England’s Clinical Expert Group for Lung Cancer; Clinical Director of the East Midlands Cancer Alliance and Honorary Professor in the School of Medicine at the University of Nottingham.
The IDEAL study (Artificial Intelligence and Big Data for Early Lung Cancer Diagnosis) has been funded by the National Institute for Health Research (NIHR) to use artificial intelligence (AI) to improve the accuracy of prediction of malignancy. Optellum was a key partner in this study and developed a risk prediction model: the Lung Cancer Prediction Convolutional Neural Network (LCP-CNN). This tool was trained using data from the US National Lung Screening Trial (NLST), marked up for machine learning applications by Optellum, under guidance from experienced thoracic radiologists at Oxford University Hospitals NHS Foundation Trust (OUH).
The LCP-CNN analyses parts of a CT scan around pulmonary nodule and provides a score from 0 to 100 for that nodule. Like the Brock model, a higher score indicates a higher chance of malignancy.
It was important to validate the model on UK clinical data because of differences between the NLST data and current UK lung nodule patients. The IDEAL study used modern data from three UK patient populations (Oxford, Leeds and Nottingham) for two purposes:
• To assess the diagnostic performance of the LCP-CNN compared to the known diagnoses in the IDEAL data set
• To compare the new model with the Brock model
The result was that the LCP-CNN performed well using the IDEAL data and significantly better than the Brock model, improving on sensitivity and specificity of diagnoses. The tool gave only one false negative cancer diagnosis, compared to six produced by the Brock model. This paper also shows that low-scoring nodules can be “ruled out” of further follow-up, potentially meaning that over 24% of nodules do not need to be followed up without missing any cancer nodules.
Instead of assigning indeterminate scores to many incidentally-detected nodules, leaving many patients undergoing long follow-up and potentially delaying life-saving cancer treatment, the LCP-CNN was shown to help clarify the patient stratification, sending more patients correctly into low-risk and high-risk groups. This successful result could have a huge impact on the efficiency of nodule management, timely diagnosis and quality of life for lung cancer patients.
Jérôme Declerck, VP of R&D at Optellum commented:
“This is an exciting time for clinical applications companies like ours, Artificial Intelligence has huge potential to assist diagnosis. We have successfully proved the clinical efficacy of the method. The next step is to confirm these results into successful clinical practice and workflow improvements.”
Further reading: Read the paper in Thorax (BMJ)