Automatic Lung Nodule Segmentation and Intra-Nodular Heterogeneity Image Generation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate lung nodule segmentation on computed tomography (CT) is difficult due to variations in nodule appearance.
• Existing methods often fail to capture intra-nodular heterogeneity, limiting their diagnostic utility.

Purpose of the Study:

• To develop an end-to-end automated system for segmenting multiple lung nodule types.
• To generate intra-nodular heterogeneity images to aid in radiological diagnosis.

Main Methods:

• A hybrid loss function was employed within a Faster R-CNN framework, incorporating generalized intersection over union loss in a generative adversarial network.
• The model was trained on a large dataset combining the Lung Image Database Consortium (LIDC) collection and data from five hospitals.

Main Results:

• The proposed model achieved an average Dice coefficient (DC) of 82.05% compared to manual segmentation by radiologists.
• It demonstrated comparable segmentation performance to established models like U-net and nnU-net.
• Generated intra-nodular heterogeneity images were found to be vivid, valid, and beneficial for diagnosis.

Conclusions:

• The developed method offers a fully automated solution for lung nodule segmentation, reducing human interaction and pre-processing needs.
• The generated intra-nodular heterogeneity images show promise in facilitating lung nodule diagnosis in clinical practice.