CoreFormer high fidelity pulmonary nodule segmentation with structural core priors and geodesic implicit fields.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate pulmonary nodule segmentation in chest CT is crucial for lung cancer diagnosis and treatment.
• Existing methods struggle with fragmented masks and topological inconsistencies due to image noise and anatomical variations.

Purpose of the Study:

• To introduce CoreFormer, a novel segmentation framework for robust and high-fidelity pulmonary nodule delineation.
• To improve boundary accuracy and topological fidelity in pulmonary nodule segmentation.

Main Methods:

• CoreFormer utilizes a Swin Transformer backbone and a dual-branch decoder with Structural Core Predictor and Context-Aware Shape Decoder.
• The framework employs structural core anchoring and geodesic shape decoding for continuous boundary generation.
• Feature Manifold Regularization enhances discriminative feature learning.

Main Results:

• CoreFormer achieved state-of-the-art performance in boundary accuracy and topological fidelity across four public datasets (LIDC-IDRI, LNDb, Tianchi-Lung, NSCLC-Radiomics).
• The method demonstrated robust and high-fidelity segmentation of pulmonary nodules.

Conclusions:

• CoreFormer offers a significant advancement in pulmonary nodule segmentation technology.
• The framework's approach to core anchoring and geodesic decoding provides superior topological consistency and boundary definition.