Efficient Multi-Organ Segmentation From 3D Abdominal CT Images With Lightweight Network and Knowledge Distillation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate segmentation of abdominal organs in Computed Tomography (CT) images is crucial for clinical applications like diagnosis and treatment planning.
• Existing 3D Convolutional Neural Networks (CNNs) for segmentation are computationally expensive due to large model sizes and high floating-point operations (FLOPs).
• The large data size of 3D CT volumes exacerbates the computational burden, limiting the clinical applicability of current segmentation methods.

Purpose of the Study:

• To develop a computationally efficient framework for segmenting multiple abdominal organs from 3D CT volumes.
• To reduce the model size and computational cost of 3D medical image segmentation.
• To improve the segmentation accuracy of lightweight models using knowledge distillation.

Main Methods:

• Proposed a novel lightweight 3D Convolutional Neural Network named LCOV-Net for efficient medical image segmentation.
• Introduced two knowledge distillation (KD) modules: Class-Affinity KD and Multi-Scale KD, to transfer knowledge from a heavy-weight model to LCOV-Net.
• Evaluated the framework on two public abdominal CT datasets for multi-organ segmentation.

Main Results:

• LCOV-Net demonstrated superior performance over existing lightweight 3D segmentation models in terms of both computational cost and accuracy.
• The proposed KD strategy significantly enhanced the performance of LCOV-Net, outperforming existing KD methods.
• The combined LCOV-Net and KD framework achieved better segmentation results than the state-of-the-art 3D nnU-Net, utilizing only one-fifth of its parameters.

Conclusions:

• The proposed lightweight network and knowledge distillation framework offer an effective and efficient solution for abdominal organ segmentation in 3D CT images.
• This approach addresses the limitations of high computational cost associated with traditional 3D CNNs, making it more suitable for clinical deployment.
• The study highlights the potential of combining lightweight architectures with advanced knowledge distillation techniques for advancing medical image analysis.