Keywords:
Artificial Intelligence, Lung, Oncology, CAD, CT, Neural networks, Cancer
Authors:
J. Murchison, G. Ritchie, D. Senyszak, E. J. R. Van Beek; Edinburgh/UK
DOI:
10.26044/ecr2019/C-3686
Aims and objectives
Despite lung cancer preventive strategies,
lung cancer remains the third highest cause of cancers worldwide [1] with a rising incidence of the disease [2].
In addition,
lung cancer is the commonest cause of cancer related deaths [3] accounting for around 1.7 million annual deaths globally.
This high mortality rate is in part due to the fact that lung cancer is often diagnosed at an advanced stage of disease.
The results of the National Lung Screening Trial (NLST) showed that early detection of lung cancer is possible using low-dose CT in a high-risk population and that this is associated with a decrease in both lung cancer related and overall mortality.
This has led to the approval of lung cancer screening in the USA [4].
The benefits of lung cancer screening and early detection of lung cancer are also supported by the findings of the Benelux NELSON trial.
Lung cancer is ideally diagnosed by histopathological confirmation on a tissue sample.
However,
the diagnostic process usually begins with detection of pulmonary nodules or masses,
usually through medical imaging.
Pulmonary nodules are very common and most are benign,
however benign and malignant nodules can have identical appearances,
so all should be flagged up as potential cancers.
The biggest challenges when it comes to pulmonary nodule detection on CT are acceptable sensitivity levels and reading times.
The importance of high sensitivity for pulmonary nodule detection is underscored by the fact that many failures in lung cancer diagnosis are due to errors of detection rather than interpretation [5,6].
Over the last two decades a substantial number of studies [7,8,9] have evaluated the performance of (sub-specialist) radiologists for the specific task of detecting pulmonary nodules and have shown that there is room for improvement.
In addition,
pulmonary nodule guidelines recommend the use of different cut-off levels for nodule size and/or volume and volume doubling time as metrics to assess nodule size and growth [10-16].
Most recently,
there seems to be consensus that semi-automated volume assessment gives the most robust assessment for lung nodules and is most helpful for determining growth during follow up.
[14-16] Apart from the measurement of size and/or volume,
another important parameter to consider is the makeup of pulmonary nodules (solid vs sub-solid),
as their malignant potential is significantly different.
[17] In spite of both these size/volume and composition management guidelines,
many hospitals currently do not have the the tools to perform these measurements in a timely and accurate manner,
both due to a lack of software and due to a lack of suitably trained radiologists.
Therefore,
software aided detection and classification of lung nodules would be a welcome addition to the radiologist’s diagnostic arsenal and could facilitate the roll-out of CT lung cancer screening as has been advocated.
[18].
The objective of this study was to evaluate the clinical performance of a Deep Learning computer assisted diagnosis system (CAD) for lung nodule segmentation on CT Chest.