ICNoduleNet: Enhancing Pulmonary Nodule Detection Performance on Sharp Kernel CT Imaging.

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Thoracic computed tomography (CT) is crucial for pulmonary nodule detection.
• Reconstruction kernel selection critically influences the performance of computer-aided detection (CAD) systems.
• The impact of kernel selection on CAD performance has been largely overlooked.

Purpose of the Study:

• To introduce a novel dataset for evaluating kernel impact on pulmonary nodule detection.
• To propose a new deep learning model for improving pulmonary nodule detection on sharp kernel CT images.
• To address the performance gap between smooth and sharp kernel reconstructions in nodule detection.

Main Methods:

• Development of the Reconstruction Kernel Imaging for Pulmonary Nodule Detection (RKPN) dataset with paired smooth (B31f) and sharp (B60f) kernel images.
• Proposal of ICNoduleNet, an image conversion-based detector utilizing a lightweight 3D slice-channel converter (LSCC) module.
• Extensive experiments to quantify detector performance differences and validate ICNoduleNet's effectiveness.

Main Results:

• Mainstream pulmonary nodule detectors exhibit superior performance on smooth kernel images compared to sharp kernel images.
• ICNoduleNet effectively converts sharp kernel images to a smooth kernel representation.
• ICNoduleNet achieves comparable or superior detection performance to baseline methods using smooth kernel images, while inputting sharp kernel images.

Conclusions:

• Reconstruction kernel selection is a critical factor in pulmonary nodule detection performance.
• The proposed RKPN dataset facilitates quantitative analysis of kernel-based imaging effects.
• ICNoduleNet offers a viable solution for enhancing pulmonary nodule detection accuracy on sharp kernel CT images.