Multi-scale and multi-view network for lung tumor segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Radiology

Background:

• Accurate lung tumor segmentation is crucial for lung cancer diagnosis and treatment planning.
• Segmentation is challenging due to variations in tumor size, shape, and contrast.
• Existing methods struggle with small 3D lung tumors.

Purpose of the Study:

• To introduce MSMV-Net, a novel deep learning architecture for enhanced lung tumor segmentation.
• To address limitations in segmenting small 3D lung tumors using multi-scale and multi-view learning.
• To improve the accuracy of lung tumor segmentation in computed tomography (CT) images.

Main Methods:

• Developed MSMV-Net, integrating multi-scale multi-view (MSMV) learning modules.
• Incorporated multi-scale uncertainty-based deep supervision (MUDS) for enhanced performance.
• Utilized CT images for segmentation, focusing on small 3D lung tumors.

Main Results:

• MSMV-Net achieved state-of-the-art performance on benchmark datasets.
• Achieved a global Dice score of 55.60% on the LUNA dataset.
• Achieved a global Dice score of 59.94% on the MSD dataset, with ablation studies confirming accuracy enhancement.

Conclusions:

• MSMV-Net demonstrates superior performance in lung tumor segmentation compared to existing methods.
• The integration of MSMV learning and MUDS effectively addresses challenges in segmenting small 3D lung tumors.
• The proposed method shows significant potential for improving lung cancer diagnosis and treatment planning.