LSW-Net: Lightweight Deep Neural Network Based on Small-World properties for Spine MR Image Segmentation

Abstract

Background: Segmenting spinal tissues from MR images is important for automatic image analysis. Deep neural network-based segmentation methods are efficient, yet have high computational costs.

Purpose: To design a lightweight model based on small-world properties (LSW-Net) to segment spinal MR images, suitable for low-computing-power embedded devices.

Study type: Retrospective.

Population: A total of 386 subjects (2948 images) from two independent sources. Dataset I: 214 subjects/779 images, all for disk degeneration screening, 147 had disk degeneration, 52 had herniated disc. Dataset II: 172 subjects/2169 images, 142 patients with vertebral degeneration, 163 patients with disc degeneration. 70% images in each dataset for training, 20% for validation, and 10% for testing.

Field strength/sequence: T1- and T2-weighted turbo spin echo sequences at 3 T.

Assessment: Segmentation performance of LSW-Net was compared with four mainstream (including U-net and U-net++) and five lightweight models using five radiologists' manual segmentations (vertebrae, disks, spinal fluid) as reference standard. LSW-Net was also deployed on NVIDIA Jetson nano to compare the pixels number in segmented vertebrae and disks.

Statistical tests: All models were evaluated with accuracy, precision, Dice similarity coefficient (DSC), and area under the receiver operating characteristic (AUC). Pixel numbers segmented by LSW-Net on the embedded device were compared with manual segmentation using paired t-tests, with P < 0.05 indicating significance.

Results: LSW-Net had 98.5% fewer parameters than U-net but achieved similar accuracy in both datasets (dataset I: DSC 0.84 vs. 0.87, AUC 0.92 vs. 0.94; dataset II: DSC 0.82 vs. 0.82, AUC 0.88 vs. 0.88). LSW-Net showed no significant differences in pixel numbers for vertebrae (dataset I: 5893.49 vs. 5752.61, P = 0.21; dataset II: 5073.42 vs. 5137.12, P = 0.56) and disks (dataset I: 1513.07 vs. 1535.69, P = 0.42; dataset II: 1049.74 vs. 1087.88, P = 0.24) segmentation on an embedded device compared to manual segmentation.

Data conclusion: Proposed LSW-Net achieves high accuracy with fewer parameters than U-net and can be deployed on embedded device, facilitating wider application.

Evidence level: 2.

Technical efficacy: 1.

Keywords: lightweight deep neural network; medical image segmentation; small-world property; spine MR image.