Lung cancer is the leading cause of cancer deaths worldwide and a diagnosis of the condition usually means a relatively poor outlook for patients. The cancer is usually only detected in the latter stages of the disease when treatment options are limited. In the early stages, surgery may be an option to remove a tumor or radiation therapy can be effective. If the cancer has spread, chemotherapy is more appropriate.
Determining whether tumors are benign or malignant and assessing the severity of the disease accurately is vital for determining the most appropriate treatment; however, diagnosing the cancer and its severity is a visual process, so the diagnosis and treatment is only based on what can be observed by radiologists. CT images of lung cancer patients can be interpreted differently, which can lead to different treatment decisions being made about patients by different oncology centers.
A team of researchers at Stanford University set out to develop a neural network that could be used to analyze the location, size, extent, and shape of lung lesions to provide a more consistent, accurate, and fast diagnosis and to more accurately assess the severity of lung cancer to guide treatment decisions.
The team of radiologists, data scientists, electrical engineers, and experts in biomedical informatics was assembled to create the neural network – LungNet – which was trained on CT scans of 4 cohorts of patients with small cell lung cancer (NSCLC). NSCLC is the most common form of lung cancer, accounting for 85% of lung cancer diagnoses. Each of the independent cohorts consisted of images relating to several hundred patients rom four different oncology centers.
The team was led by Olivier Gevaert, Assistant Professor of Medicine in Biomedical Informatics Research at Stanford University who is a specialist in machine learning using multi-scale biomedical data. “Quantitative image analysis has demonstrated that radiological images, such as CT scans of patients with lung cancer, contain more minable information than what is observed by radiologists,” said Gevaert. “Using CT image datasets from several different oncology clinics, we set out to determine whether our neural network could be trained to accurately and reproducibly analyze the scans and deliver consistent, useful clinical information.”
The machine learning system was able to accurately differentiate between benign and malignant nodules in the lungs, stratify nodules regarding cancer progression, and accurately predict overall survival for patients in each of the four patient groups.
The researchers believe their machine learning system can be used as a tool to help diagnose lung cancer and stratify patients into low, medium, and high risk groups, which will allow patients in the higher risk groups to be identified more rapidly and sent for intensified treatment, while reducing unnecessary treatments for patients in low risk groups.
You can read more about the machine learning system in the paper – A shallow convolutional neural network predicts prognosis of lung cancer patients in multi-institutional computed tomography image datasets – which was recently published in the journal Nature Machine Intelligence. DOI: 10.1038/s42256-020-0173-6