A novel approach to early detection of cancerous lung nodules
Lung cancer is the leading cause of cancer related deaths in the world. As per the 2018 ‘State of Lung Cancer’ report by American Lung association, every two and a half minutes someone in the United States will be diagnosed with lung cancer, with an estimated 234,030 new cases in the United States this year. The toll of lung cancer varies across nations. In India too, it is the one of the leading causes of cancer related deaths.
A delayed lung cancer diagnosis is associated with a poor prognosis. About 60 per cent of patients dies within 1 year of diagnosis. Surgical resection of an early lung cancer has a favourable prognosis. After resection of a first stage bronchogenic carcinoma, the 5-year survival rate is 80 per cent to 90 per cent. Therefore, significant efforts targeted towards prevention and early detection are very crucial.
The challenge of lung cancer detection
Early detection of lung nodules in patients at risk for lung cancer can be achieved through Chest Computed Tomography (CT) screening. This has shown to have improved survival rate and such programs have been implemented in clinical settings.
The arbitrariness of shape, size and texture of lung nodules is a challenge when detecting the presence of nodules. Radiologists analyze the chest CT scans with their expertise in human anatomy and experience in looking for structures of interest. They identify nodules in the image and annotate them for further evaluation. Rapid development of CT scanning technologies, increasing demand, more than hundreds of images per exam, other pathology distracts in the image are some of the challenges faced by the radiologists in this process. Moreover, repetitive analysis and precise detection of nodules consumes a huge amount of radiologist’s effort and time. Often this leads to work pressure and heavy workload resulting in many nodules being missed in clinical practice.
This has led to more focus being given to Computer Aided Detection (CAD) and diagnosis of lung nodules. CAD systems can be very helpful for achieving a lung cancer screening workflow, which is efficient as well as cost-effective.
A single CT scan of a patient can contain 300 to 400 axial images. Looking for nodules in these scans requires a search through approximately 300 axial sections, each composed of over 260,000 pixels. A 5-mm lesion occupies a small fraction of the image area and occurs within a background of highly complex lung tissue. In conventional CAD systems, the whole process of detection involves generation of candidate nodules and classification of such nodules based on features which differentiates a true nodule from a non-nodule. Since the features are handcrafted, the accuracy and sensitivity of such systems are low.
Using deep learning to achieve real time diagnosis
Often described as a subset of machine learning, deep learning deals with algorithms inspired by the structure and function of the human brain, called artificial neural networks (ANN). It involves a computational model which is made to learn a problem, by training it on huge amounts of data associated with that problem. By getting trained on data, the model or network learns a set of parameters or weights. The fully trained model can be deployed to make inferences using the learned weights to do operations like classification, detection etc. --in simple terms the model mimics the brain. Deep learning algorithms called convolutional neural networks (CNN) are now used extensively in solving computer vision problems.
Nodule detection from CT scans using deep learning can give better performance, accuracy and sensitivity than other conventional detection methods. This in turn helps doctors to minimize the time spent on the detailed visual inspection of the CT images. Their valuable time can be used for communicating with patients, planning the treatment or for further research work. Moreover, detecting the presence of nodules early, leads to better cancer diagnosis and the survival rate can be improved which makes the whole process really rewarding.
Principal Architect, QuEST Global