Performance analysis of a computer-aided detection system for lung nodules in CT at different slice thicknesses

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Computed tomography (CT) slice thickness impacts lung nodule detection.
• Computer-aided detection (CAD) systems are crucial for identifying lung nodules.
• Utilizing diverse CT slice thicknesses in training data presents challenges.

Purpose of the Study:

• To evaluate CAD system performance for lung nodules across varying CT slice thicknesses.
• To compare three distinct training methodologies for non-homogeneous CT data.
• To determine the optimal approach for training CAD systems with varied slice thickness data.

Main Methods:

• Assessed CAD performance using CT datasets with 1.25 mm and 2.5 mm slice thicknesses.
• Compared three training strategies: aggregate, homogeneous subset, and common thickness resampling.
• Utilized data from the Lung Nodule Analysis 2016 dataset.

Main Results:

• CAD performance at 2.5 mm slice thickness was comparable to 1.25 mm and superior to higher thicknesses.
• Resampling all training and testing data to a common thickness (2.5 mm) yielded the best results.
• The common thickness resampling method improved accuracy, reduced memory usage, and decreased computational time.

Conclusions:

• CT slice thickness significantly influences CAD system performance for lung nodule detection.
• Resampling CT data to a consistent slice thickness is an effective strategy for optimizing CAD systems.
• Findings have implications for CT acquisition, processing, and data storage protocols.