Empirical Driven Automatic Detection of Lobulation Imaging Signs in Lung CT

Area of Science:

• Radiology
• Medical Imaging
• Computer Science

Background:

• Computer-aided detection (CAD) aids radiologists in diagnosing lung diseases.
• Lobulation detection is an underexplored area in CAD due to its complex nature.
• Existing state-of-the-art methods struggle with accurate lobulation detection.

Purpose of the Study:

• To develop an effective computer-aided detection framework for lung lobulation.
• To investigate classical methods for lobulation detection, focusing on undulated characteristics.
• To improve the accuracy and efficiency of diagnosing lung diseases through lobulation identification.

Main Methods:

• A sliding window-based framework was designed for lobulation detection.
• Three categories of lobulation classification algorithms were investigated: template matching, feature-based classifiers, and bending energy.
• Experiments were conducted on the LISS database to evaluate detection algorithms.

Main Results:

• The algorithm combining global context features and Bag of Visual Words (BOF) encoding achieved the best overall performance with an F1 score of 0.1009.
• Bending energy was found effective in reducing false positives.
• Applying bending energy after LIOP-LBP feature extraction improved the F1 score from 0.0599 to 0.0643 and reduced average positive detections per image from 30 to 22.

Conclusions:

• This work represents the first direct lobulation detection study and the initial application of bending energy to lobulation analysis.
• The developed framework and investigated methods offer a promising approach for improving lung disease diagnosis.
• Further research can build upon these findings to refine CAD systems for lobulation detection.