Can Artificial Intelligence Tell Normal Lung Anatomy from Diseased Lung Tissue?
Linking two neural networks outperforms other methods in detecting pulmonary fissures.
Medical imaging technologies continue to improve by leaps and bounds, and the use of artificial intelligence to make sense of CT scans holds the promise of increasing the speed and accuracy of 3D image interpretation. For pulmonary disease, CT scans are the standard way to visualize lung structure, but existing methods to delineate the fissures between lung lobes do a poor job of handling the natural variability in lobar anatomy and often mistake normal tissue for diseased tissue. Automating the process of mapping anatomical structures, called segmentation, usually requires software that can interpret both local data and the larger context, i.e., the fine structure of a fissure in its global environment of the lung. This takes enormous computing power.
This summer in the journal IEEE Transactions on Medical Imaging, researchers from the University of Iowa and the University of Wisconsin-Madison reported on FissureNet, an AI method to dramatically reduce the processing power required to accurately render both the local fissure environment and its surroundings. They linked two existing convolutional neural networks, training one to define a fissure region of interest and the other to detect the precise fissure location within that region.
The researchers tested FissureNet with data from people with chronic obstructive pulmonary disease and lung cancer whose scans had been segmented by human experts. They also compared FissureNet with two rule-based fissure computer detection methods and one learning-based method. FissureNet did as well as the other learning method and both outperformed the rule-based methods. It worked equally well on data produced by different scanners and scanning modalities.
FissureNet is designed only to detect fissures and does not produce a complete lobar segmentation, but the researchers wrote that this final step will be easy to complete with “straightforward post-processing.”