Lung cancer early detection is made difficult by the small size of pulmonary nodules, which cannot be detected by regular X-rays, and by the long time needed to classify the nodules seen on CT scans in thousands of new patients every day. Computer-aided lung nodule classification offers support to radiologists by providing automatic detection and analysis to speed and improve their manual observations.
Accuracy of results is very high, though it depends on the visual similarities of nodules: the more dissimilar they are, the more reliable the lung nodule classification will be. This will also enable to keep false positives at a very low level, though of course the main concern of the study is to avoid false negatives, which would prevent the timely treatment of the disease with potentially fatal consequences. In this area, the system compellingly outperforms traditional methods. Its main breakthrough is that deep learning features also take into account the association between the different morphologic findings in the lung like presence or absence of lobulation, coarse spiculation and so on.
A very promising future direction of the research will require to integrate in the system an automatic detection of nodules with no previous use of human involvement such as radiologist-constructed lung nodule outlines.
* Source: cancer.org