Semi-automatic classification of textures in thoracic CT scans

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Pulmonary Medicine

Background:

• Accurate diagnosis of interstitial lung disease (ILD) relies on identifying textural patterns in lung parenchyma on computed tomography (CT) scans.
• Automated and interactive methods can aid in classifying these complex textural patterns.

Purpose of the Study:

• To develop and compare one automatic and two interactive protocols for classifying normal and seven types of abnormal lung textures in CT scans.
• To evaluate the efficiency of these protocols by measuring the percentage of volumes of interest (VOIs) requiring observer relabeling.

Main Methods:

• Lungs were segmented and subdivided into VOIs using clustering.
• An extra-trees classifier was employed for texture classification.
• Protocols included automatic classification, and two interactive methods (with and without prior training data).

Main Results:

• The automatic protocol required observers to relabel 58% of VOIs.
• Interactive protocols significantly reduced relabeling: 21% overall without prior data, and 20% overall with prior data.
• Observer effort decreased substantially after the initial slices in interactive protocols.

Conclusions:

• Interactive classification protocols, especially with prior training data, are more efficient than automatic methods for classifying lung textures on CT.
• These methods can reduce observer workload in diagnosing interstitial lung diseases.