Deep learning for lung disease segmentation on CT: Which reconstruction kernel should be used?

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Deep learning models are increasingly used for medical image analysis.
• Chest computed tomography (CT) reconstruction kernels can influence image texture and detail.
• The choice of kernel may impact the performance of AI models for lung disease segmentation.

Purpose of the Study:

• To compare the effectiveness of single versus combined reconstruction kernels for training deep learning models.
• To evaluate model performance for segmenting diffuse lung diseases, including COVID-19 pneumonia and interstitial lung disease (ILD).

Main Methods:

• U-Net architecture trained on annotated CT datasets (COVID-19 and ILD).
• Images reconstructed using a lung kernel (LK) and a mediastinal kernel (MK).
• Models trained on LK only, MK only, or LK+MK images; performance compared using Dice Similarity Scores (DSC).

Main Results:

• Models trained on a single kernel performed best on images reconstructed with the same kernel.
• Models trained on both LK and MK demonstrated improved or comparable performance across both kernel types.
• Combined kernel training yielded higher DSC for both COVID-19 and ILD segmentation compared to single-kernel training.

Conclusions:

• Reconstruction kernel choice significantly impacts deep learning model performance for lung disease segmentation.
• Training deep learning models with both LK and MK images enhances segmentation accuracy for diffuse lung diseases on CT.