Few-shot segmentation framework for lung nodules via an optimized active contour model.

Area of Science:

• Medical Imaging Analysis
• Computational Pathology
• Artificial Intelligence in Healthcare

Background:

• Accurate lung nodule segmentation is vital for lung cancer diagnosis and treatment.
• Challenges in segmentation arise from the visual similarity between lung nodules and surrounding tissues.

Purpose of the Study:

• To integrate deep learning and active contour models to overcome individual limitations.
• To develop a robust lung nodule segmentation framework with improved accuracy and reduced data dependency.

Main Methods:

• Proposed a few-shot segmentation framework combining deep neural networks and active contour models.
• Introduced heat kernel convolutions and high-order total variation into the active contour model.
• Utilized presegmentation from a few-shot deep neural network as initial contours for the active contour model.

Main Results:

• The proposed method demonstrated outstanding segmentation performance on clinical CT images and the LIDC dataset.
• Achieved superior results compared to state-of-the-art methods based on visual and quantitative evaluations.
• Validated effectiveness using data from two hospitals and a public dataset.

Conclusions:

• The framework leverages few-shot learning outputs as prior information, eliminating manual initial contour selection.
• Offers mathematical interpretability for deep learning models, reducing dependence on extensive training data.
• Enhances the feasibility of accurate lung nodule segmentation in clinical settings.