Currently, lung cancer has become one of the most deadly cancers. In contrast to the steady increase in survival for most cancers, advances have been slow for lung cancer. Commonly, the five-year survival rate for lung cancer patients is merely 16%, which will rise to 52% if lung cancer is diagnosed at early stages. However, it will decrease to below 4% if cancer spread occurs. Therefore, it is of crucial importance to detect lung cancer at early stages for prolonging patient’s life. Liu et al. from Beihang University have proposed one computer aided detection method based on artificial neural network for lung cancer detection, which will soon appear in the 7th issue of SCIENCE CHINA Information Sciences in 2017.
In clinical practice, Computer Tomography (CT) can capture fine-grained details for both lung nodules and surround structures, acting as the golden standard for diagnosis. However, the high sensitivity of CT imaging also leads to huge data and complex ambiguities, which makes it hard for radiologists for distinguishing pathological structures from healthy. In recent years, Computer Aided Detection (CADe) system has developed rapidly and shown great potential in diagnosis assistance. Detection for lung nodules is an obvious guidance for lung cancer diagnosis and treatment. However, it is hard to assess lung nodules due to various nodule appearances, minor differences between nodule and healthy structures, as well as the influence by vessel and other tissues around nodules.
Inspired by the prior works, this article presents an artificial neural network based approach to the extraction of lung nodules from chest CTs. The pipeline and the ANN architecture can be found in Fig.1 and Fig.2, separately. Different from classical methods, we focus on the inner structures of nodule voxels and apply ANN to generalize these characteristics. We are working in 3D space consisting of only voxels instead of processing slice by slice in CT volume. Our method can be easily integrated into existing CADe systems and rapidly accommodate and process new data streams with few human interactions. Meanwhile, we propose a novel voting method based on geometrical and statistical features to better extract initial candidate regions while suppressing ambiguous structures. Finally, we have proposed a nodule detection approach with multiple trained ANNs based on 3D massive sampling of candidate voxels instead of user-specified features with a goal to reduce various false positives.
Please refer to Liu X, Hou F, Qin H, et al. A CADe system for nodule detection in thoracic CT images based on artificial neural network[J]. Science China Information Sciences, 2017, 60(7):072106. for details, which can be found under http://engine.
|DevilsRishon Le Zion||9||7||1||1||22||68||22||+46|
|Dragons 2Ness Ziona||5||3||2||0||9||36||19||+17|
|Dragons 3Ness Ziona||5||3||2||0||8||14||19||-5|
|Jet TurtlesBat Yam||5||0||5||0||0||9||41||-32|
|MonstersRishon Le Zion||6||3||2||1||10||34||18||+16|
|Silver Fox 2Yehud||6||3||3||0||8||35||36||-1|
|LegionRishon Le Zion||6||1||5||0||3||18||30||-12|
|Silver Fox 2Yehud||3||2||1||0||6||18||13||+5|
|Dragons 2Ness Ziona||1||1||0||0||3||15||4||+11|
|MonstersRishon Le Zion||2||1||1||0||3||5||6||-1|
|LegionRishon Le Zion||2||1||1||0||3||6||8||-2|
|Jet TurtlesBat Yam||1||0||1||0||0||3||4||-1|
|Dragons 3Ness Ziona||3||0||3||0||0||6||18||-12|