Automated lung tumor segmentation robust to various tumor sizes using a consistency learning-based multi-scale...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Automatic lung tumor segmentation is difficult due to significant variations in tumor size (less than 1cm to over 7cm).
• Tumor size impacts segmentation accuracy, often leading to under- or over-segmentation.

Purpose of the Study:

• To develop and evaluate a novel consistency learning-based multi-scale dual-attention network (CL-MSDA-Net) for accurate lung tumor segmentation.
• To address challenges posed by diverse lung tumor sizes in automated segmentation.

Main Methods:

• A size-invariant patch generation technique normalizes tumor-to-structure ratios.
• A dual-branch network utilizes consistency learning, sharing weights and employing a consistency loss function.
• Each branch incorporates a multi-scale dual-attention module to enhance feature learning across scales and attention mechanisms.

Main Results:

• On hospital datasets, CL-MSDA-Net achieved an F1-score of 80.49%, outperforming U-Net variants by up to 3.91%.
• On NSCLC-Radiomics datasets, CL-MSDA-Net achieved an F1-score of 71.7%, outperforming U-Net variants by up to 3.66%.
• The network demonstrated superior performance across various tumor sizes, with notable improvements for smaller tumors.

Conclusions:

• CL-MSDA-Net effectively enhances lung tumor segmentation accuracy across all tumor sizes.
• The proposed network shows significant potential for improving diagnostic and treatment planning in lung cancer.