Lung Cancer Prediction

Unmet need

Lung cancer is the biggest cause of cancer death in the UK and worldwide, with £307M/year cost to NHS England. Earlier diagnosis is critical for increasing survival, and the current diagnostic pathways can be improved. Randomised controlled trials show that screening programmes can reduce mortality by 20-26%, and detect co-morbid disease and has led to the establishment of a new £70M NHS England lung cancer screening programme.

Lung cancer remains the leading cause of cancer-related death in the United States and worldwide. In the United States alone, an estimated 234,000 cases were diagnosed in 2018 and despite recent progress in immunotherapy and other treatment modalities, the five-year survival rate is 18.6%.

Early diagnosis can markedly improve outcomes: survival for patients with stage IA1 non-small cell cancer is 92%. There are two principal mechanisms by which lung cancer may be diagnosed this early. The first is through screening, using low-dose computer tomography (LDCT) which was shown to reduce lung cancer deaths by 20% in the USA National Lung Screening Trial (NLST), and by 26% in the European NELSON trial.

The second mechanism is the detection of cancer as incidental findings in patients undergoing imaging for unrelated reasons. Indeterminate Pulmonary Nodules (IPNs) are reported as incidental findings in approximately 30% of chest CTs and it has been estimated that 1.57 million patients with pulmonary nodules are identified in this way every year in the USA. 

Project aim

Optellum has developed a digital biomarker to predict a lung nodules probability of malignancy using AI. However, the indications for use of the first AI are limited to patients with indeterminate pulmonary solid and semi-solid nodules between 5 and 30mm in diameter. Not all patient groups, including those already with a history of cancer can participate.

The NCIMI project is collecting data for Optellum to be able to study the potential to extend its AI to such patients, potentially extending its coverage to many more patients, and increasing its utility to pulmonologists and radiologists reading these difficult cases.

Objectives

1. Collect a longitudinal CT dataset comprising 1000 patients with benign and malignant nodules, with at least 1/3 of patients with malignant nodules. 
2. Collect a longitudinal CT dataset comprising 1000 patients with benign and malignant nodules of any type from patients with a history of cancer within the last 5 years. 
3. Evaluate the baseline performance of the current LCP-CNN on the above datasets.
4. Update the LCP-CNN (Lung Cancer Prediction Convolutional Neural Network) to include GGOs as a nodule sub-types and evaluate the performance.
5. Update the LCP-CNN to include patients with a history of cancer and evaluate its performance.