This project aims to develop and evaluate the impact of a multivariable lung nodule malignancy predictor for the detection of lung nodules, which may go on to develop into cancerous growth. We know that the risk of these nodules is not uniform, and some have characteristics that may increase the risk for cancer development. To date, standard practice for assessing malignancy risk for incidental lung nodules is either the subjective risk assessment of an experienced thoracic radiologist or performed using a linear regression model referred to as the Brock model. The Brock model was developed from participants enrolled in the Pan-Canadian Early Detection of Lung Cancer Study (Pan-Canadian Detection of Lung Cancer Study), has been validated in lung cancer screening and clinical populations and is recommended by the British Thoracic Society guidelines for pulmonary nodules.
New advances in the field of artificial intelligence (AI) have resulted in alternative malignancy prediction models that have shown to outperform the Brock model in lung cancer screening cohorts. Development and validation of similar prediction models for incidental lung nodules in clinical practice have not been established. The main reason for this is that it is challenging to build relevant cohorts due to an absence of well indexed datasets at local institutes together with a relatively low incidence of lung cancer in a single institute. This PICTURES exemplar can meet these limitations and provide a well-defined and stratified cohort including both proven benign and malignant incidental lung nodules.
The output of this study will be two-fold:
- A multivariable predictor that uses logistic regression to estimate likelihood of lung nodule malignancy from multiple variables, such as age, sex, family history, and nodule characteristics including risk scores derived from the pixel data in CT scans.
- An imaging based predictor that uses deep-learning techniques such as convolutional neural networks to predict the probability of lung nodules on a CT scan becoming cancerous, trained on CT pixel data linked to patient outcomes.
The multivariable predictor will be publicly accessible and function in a similar way to existing models such as the Lung Nodule Risk Calculators found here https://www.sts.org/resources/lung-nodule-resources/lung-nodule-risk-calculators. We believe our model will outperform previous models because it has been developed using routinely collected, population level data and risk scores from deep learning models trained on CT scans. The imaging based predictor will be developed by our industry partner Aidence B.V. from anonymised data shared with them for analysis in their own development environment, with a view to commercialising the imaging based predictor through integration with their existing Veye Lung Nodule device.
The primary benefits of this exemplar project will be the improvement of existing publicly available risk assessment tools and integration of new proprietary ones into the radiology pathway, to support early detection of lung cancer. On the proprietary aspects of the project, Aidence B.V. would own the IP on medical software devices developed by them but have agreed to make such available to NHS Scotland with a non-exclusive, non-transferable, commercial licence discount. This industry partnership will also contribute in-kind transfer of expert skills and knowledge transfer to develop NHS Scotland staff skills and capabilities.