Nested block self-attention multiple resolution residual network for multiorgan segmentation from CT.

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Anatomy

Background:

• Accurate multiorgan segmentation in CT scans is crucial for radiation therapy planning.
• Self-attention (SA) deep learning methods offer high accuracy but are computationally intensive, limiting their use in deep networks.
• Existing methods struggle with the computational demands of deep self-attention models for complex segmentation tasks.

Purpose of the Study:

• To develop and validate a novel, computationally fast, and memory-efficient bidirectional SA method named nested block self-attention (NBSA).
• To enable accurate multiorgan segmentation in both shallow and deep neural networks.
• To improve the efficiency and applicability of SA techniques in medical image analysis.

Main Methods:

• Developed a deep multiple resolution residual network with NBSA (MRRN-NBSA) for segmenting 18 organs in head and neck (HN) and abdomen CT scans.
• MRRN-NBSA integrates multi-resolution features with bidirectional SA for enhanced contextual feature extraction.
• Evaluated MRRN-NBSA on public datasets, comparing its performance against Unet, CCA, dual SA, and UNETR using Dice Similarity Coefficient (DSC) and statistical tests.

Main Results:

• MRRN-NBSA achieved superior average DSC scores: 0.88 for HN and 0.86 for the abdomen, outperforming current methods.
• Demonstrated significantly higher accuracy than CCA (0.845 HN, 0.727 abdomen) and other compared methods (Unet, dual SA, UNETR) (p < 0.05 after Bonferroni correction).
• MRRN-NBSA exhibited more consistent segmentation for submandibular glands (0.82 ± 0.06) compared to manual raters (0.75 ± 0.31).

Conclusions:

• The MRRN-NBSA method provides more accurate multiorgan segmentations compared to existing techniques across different public datasets.
• The developed NBSA approach offers a computationally efficient and effective solution for medical image segmentation.
• Further validation on larger clinical datasets is recommended to confirm its clinical utility.