Detection-guided deep learning-based model with spatial regularization for lung nodule segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Lung cancer is a leading cause of cancer mortality worldwide.
• Early lung nodule detection is crucial for improving patient outcomes.
• Accurate segmentation of lung nodules aids in distinguishing benign from malignant lesions, but is challenging due to variations in nodule appearance and proximity to lung tissues.

Purpose of the Study:

• To develop an accurate and reliable lung nodule segmentation method.
• To assist radiologists in improving diagnostic accuracy for lung cancer.
• To create a model that can effectively segment and classify lung nodules using deep learning.

Main Methods:

• A novel deep learning model integrating U-Net based segmentation and ResNet based classification.
• Feature combination blocks for information sharing between segmentation and classification components.
• Classification outcomes used as priors to refine segmentation size estimation and spatial regularization for enhanced precision.
• Optimal transfer learning (TL) strategy with frozen layers to address limited training datasets.

Main Results:

• The proposed model demonstrated superior accuracy in capturing target nodules compared to existing models.
• Ablation studies confirmed the positive impact of feature combination and spatial regularization.
• Transfer learning further improved performance, achieving a sensitivity of 0.885, Dice score of 0.814, Hausdorff distance of 3.188 mm, and ASSD of 0.280 mm.

Conclusions:

• The multitask model effectively performs lung nodule segmentation and classification.
• The model achieves strong performance with limited training data through transfer learning.
• The model's adaptability provides a foundation for enhancements on specific datasets.