Segmentation of lung nodules based on a refined segmentation network

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Radiology and Oncology

Background:

• Accurate lung nodule segmentation is crucial for early lung cancer detection and diagnosis.
• Challenges in segmentation include nodule heterogeneity and similarity to surrounding lung tissues.
• Deep learning models, particularly convolutional neural networks, face segmentation errors due to layered upsampling/downsampling.

Purpose of the Study:

• To develop a refined segmentation network (RS-Net) for accurate lung nodule segmentation.
• To address the limitations of existing deep learning methods in handling nodule heterogeneity and small nodule misdetection.
• To improve the precision of lung nodule segmentation for enhanced early cancer screening.

Main Methods:

• Developed a refined segmentation network (RS-Net) focusing on core region localization and gradual refinement.
• Implemented an average dice-loss function computed at the nodule level to mitigate issues from imbalanced sample sizes.
• Evaluated the network's performance on diverse lung nodule datasets.

Main Results:

• RS-Net achieved segmentation dice coefficients of 85.90% on the LIDC dataset and 81.13% on the Shanghai Chest Hospital dataset.
• The network demonstrated a stable segmentation effect across nodules of varying properties and sizes.
• The refined segmentation strategy proved effective in improving lung nodule segmentation accuracy.

Conclusions:

• The proposed RS-Net and its gradual refinement strategy significantly enhance lung nodule segmentation.
• The nodule-level average dice-loss function effectively addresses the misdetection of small nodules.
• This approach offers a promising tool for improving the accuracy and reliability of lung cancer screening.