Pulmonary nodules are small growths in the lung that can be detected, often incidentally, using a computed tomography (CT) scan. The majority of nodules are benign but they may be an early lung cancer and so warrant further investigation. However, it is challenging to discriminate a benign from a malignant nodule with a single CT scan. Currently, a number of repeat scans over a period of up to 2 years may be required to identify whether the nodule is growing, which would indicate an increased likelihood of cancer. Further tests would then be required to confirm a cancer diagnosis. Improving the ability to predict malignancy using the initial CT scan would remove a delay from the diagnostic pathway, increasing the window of opportunity for an early cancer diagnosis, when treatment is more likely to be successful.
The British Thoracic Society currently recommends the use of the Brock University model that incorporates patient clinical information and nodule characteristics to predict which nodules are likely to be malignant. A new model, the Lung Cancer Prediction Convolutional Neural Network (LCP-CNN) has been developed by Optellum Ltd, an Oxford-based company. This model uses artificial intelligence (AI), rather than logistic regression as in Brock, to extract nodule information from the CT scan alone and predict malignancy by comparing it to data from thousands of nodules where the diagnosis is known. An AI-based system can account for nodule size and other radiological factors consistently, and without requiring subjective judgement or data entry by a clinician.
The LCP-CNN model was externally validated and compared with the Brock University model as part of the IDEAL study, led by Professor Fergus Gleeson (Oxford University Hospitals NHS Foundation Trust (OUH FT)/Department of Oncology) with Optellum Ltd and clinical teams from OUH FT, Nottingham University Hospitals and the University of Nottingham (led by Professor David Baldwin), and the Leeds Teaching Hospitals NHS Trust (led by Dr Matthew Callister). The results of this analysis, published in the journal Thorax, showed that both models performed well but the LCP-CNN model discriminated benign from malignant nodules better, with fewer false negatives, potentially allowing the LCP-CNN model to be employed in the future to rule out low-risk nodules for further follow-up. The LCP-CNN model is currently undergoing rigorous testing in a representative prospective clinical environment. The aim is to implement the LCP-CNN model into a clinical system that will support clinicians, deliver improvements to patient care, and save the NHS money.
Further evidence to support the clinical validity of the LCP-CNN has also just been published in the American Journal of Respiratory and Critical Care Medicine.
This work was funded by the National Institute for Health Research’s i4i Programme, IDEAL: Artificial Intelligence and Big Data for Early Lung Cancer Diagnosis, i4i/II-LB-0716-20006.