SVNC-Net: An optimized U-Net variant with 2D convolutions for lightweight 3D spleen segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate spleen volume measurement is crucial for diagnosing splenomegaly.
• Manual spleen segmentation in Computed Tomography (CT) is time-consuming and not practical for routine clinical use.
• Automated segmentation methods, especially 3D Convolutional Neural Networks (CNNs), offer a viable alternative but face computational and memory challenges.

Purpose of the Study:

• To introduce SVNC-Net, an efficient deep learning model for 3D spleen segmentation from CT scans.
• To optimize the U-Net architecture for improved efficiency in spleen segmentation.
• To enable real-time spleen segmentation on edge devices with limited resources.

Main Methods:

• Developed SVNC-Net, a U-Net based architecture utilizing 2D convolutions and depthwise separable convolutions for efficiency.
• Evaluated SVNC-Net against established CNN models (UPerNet, EMANet, CCNet, SegNet, ShuffleNet) on two public datasets.
• Applied post-training compression techniques, including pruning and quantization, to further enhance model performance.

Main Results:

• SVNC-Net demonstrated high suitability for real-time applications and resource-constrained environments.
• Comparative analysis confirmed the efficiency and effectiveness of SVNC-Net compared to other models.
• Post-training compression significantly improved the model's compactness and inference speed.

Conclusions:

• SVNC-Net offers an efficient solution for 3D spleen segmentation in CT scans.
• The model's design addresses the limitations of traditional 3D CNNs for clinical deployment.
• This work advances the development of efficient deep learning models for 3D organ segmentation in resource-limited clinical settings.