SwinD-Net: a lightweight segmentation network for laparoscopic liver segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Surgical Technology

Background:

• Real-time image segmentation is critical for laparoscopic surgical assistance systems.
• Traditional deep learning models offer high accuracy but are computationally intensive, limiting their use in resource-constrained hospital settings.
• There is a need for efficient segmentation models that balance accuracy and computational overhead.

Purpose of the Study:

• To propose a novel, lightweight deep learning network, SwinD-Net, for real-time image segmentation in laparoscopic surgery.
• To reduce computational burden and parameter count while maintaining high segmentation accuracy.
• To validate the effectiveness of SwinD-Net on the CholecSeg8k dataset.

Main Methods:

• Developed SwinD-Net incorporating Skip connections, Depthwise separable convolutions, and Swin Transformer Blocks.
• Optimized the network by eliminating the first layer skip connection and reducing shallow feature map channels.
• Introduced Swin Transformer Blocks to capture global information and high-level semantic features.

Main Results:

• SwinD-Net achieved high accuracy on the CholecSeg8k dataset with significantly reduced computational overhead.
• The model requires only 98.82 M FLOPs and 0.52 M parameters, with an inference time of 47.49 ms/image on a CPU.
• Outperformed UNeXt in Dice metric, with 1/3 parameters, 1/22 FLOPs, and 2.4x faster inference speed.

Conclusions:

• SwinD-Net effectively reduces parameter count and computational complexity, enhancing inference speed while maintaining comparable accuracy.
• The lightweight design makes SwinD-Net suitable for real-time applications in hospitals with limited computing resources.
• The proposed network offers comprehensive improvements in accuracy, speed, and efficiency for surgical image segmentation.