Mate Denes

MSc.

Ulyssys Ltd.

Researcher and developer

NLST (Learn more about this study)

NLST-627

Jan 10, 2020

Development of lung cancer diagnostic artificial intelligence based on CT images

More than 1.2 million people die of cancer, the dreadful illness of our age in the European Union every year. Almost every fourth of the cancer cases is lung cancer, so this is the most common among cancerous diseases. Moreover, lung cancer has very bad survival rates. The likelihood of survival of lung cancer, like of other cancerous diseases, depends on the discovery stage to a great extent - whilst in the first stage 55 to 75% is the five-year survival rate in the fourth it is only 2-13%.
Therefore, the possible introduction of lung cancer screening was considered in many countries of the world. Unfortunately, x-ray-based lung cancer screening does not reveal the lesions with high enough certainty, thus CT imaging is needed. The European Union's expert group also took a stand on this in 2017 proposing the introduction of CT lung cancer screening for the risk groups in all EU countries within 4 years.

Beyond the need for screening machines, one of the main obstacles to the introduction of screening is that it is extremely expensive to process huge amounts of CT images and to make diagnoses with radiologists, moreover it is not even certain that there are enough such experts to handle this workload at all. Therefore, in recent years the need emerged to develop computerized image analysis programs that recognize cancerous lesions with a high degree of precision in a CT scan.
There have been experiments with traditional CAD systems for a long time, which detect suspicious lesions with enough accuracy, but there is a large amount of error in detecting whether they are cancerous. However, a few years ago Artificial Intelligence (AI) based computer systems achieved great breakthroughs in solving object recognition tasks that are difficult to solve with traditional algorithmic methods.

Deep Convolutional Neural Network (DCNN) is an explosively developing branch of artificial intelligence research today. This technology achieved a breakthrough in object recognition in 2014. Our project is building on the recent achievements of this field, collecting the approrpiate number of CT images, applying radiological competence for buidling the training data set and synthetizing the radiological and AI researchers efforts for buidling and training of the appropriate DCNN.