MRUNet-3D: A multi-stride residual 3D UNet for lung nodule segmentation

Area of Science:

• Medical Imaging
• Radiology
• Computer Vision

Background:

• Accurate pulmonary nodule segmentation in CT images is crucial but challenging due to nodule heterogeneity and similar visual characteristics with surrounding tissues.
• Existing methods, including UNet variants, often lack robust multi-scale feature extraction capabilities necessary for precise segmentation.

Purpose of the Study:

• To propose and evaluate a novel Multi-Stride Residual 3D UNet (MRUNet-3D) for enhanced segmentation accuracy of pulmonary nodules in CT images.
• To address the limitations of current segmentation techniques in handling nodule heterogeneity and improving the detection of small nodules.

Main Methods:

• Introduction of a Multi-Stride Residual 3D UNet (MRUNet-3D) incorporating a Multi-Stride Res2Net (MSR) block in the encoder stages.
• The MSR block replaces standard convolutional layers to extract multi-scale features at granular levels from diverse receptive fields and resolutions.
• Evaluation on the LUNA16 dataset using metrics such as Dice Similarity Coefficient and Surface Distance.

Main Results:

• The MRUNet-3D achieved a competitive Dice Similarity Coefficient of 83.47% and an average Surface Distance of 0.35 mm on the LUNA16 dataset.
• The proposed method demonstrated robustness in segmenting heterogeneous lung nodules.
• MRUNet-3D showed improved performance in segmenting small pulmonary nodules.

Conclusions:

• The proposed MRUNet-3D, with its MSR and RFIA modules, significantly enhances pulmonary nodule segmentation accuracy in CT images.
• The method effectively addresses challenges posed by nodule heterogeneity and improves the segmentation of small nodules.
• MRUNet-3D offers a promising advancement for automated lung nodule analysis in medical imaging.