Efficient knowledge distillation for liver CT segmentation using growing assistant network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning (DL) models excel in medical image segmentation but are often too large for real-time surgical applications.
• Existing DL segmentation models are computationally expensive and memory-intensive, limiting their use in liver surgery.
• Compressing DL models often leads to a performance trade-off between model size and accuracy.

Purpose of the Study:

• To develop a real-time, 3-D liver CT segmentation method suitable for liver surgery intervention.
• To address the computational and memory inefficiencies of current DL segmentation models.
• To achieve model compression without sacrificing segmentation performance.

Main Methods:

• Proposed a deep learning-based real-time 3-D liver CT segmentation method.
• Incorporated knowledge distillation (KD) to transfer knowledge from a large teacher model to a smaller student model.
• Introduced a growing teacher assistant network (GTAN) to facilitate efficient knowledge transfer between models of significantly different sizes.

Main Results:

• The student model with KD achieved a 1.2% improvement in Dice similarity coefficient (DSC) compared to the student model without KD (85.9% to 87.1%).
• The KD-enhanced student model reached performance comparable to the teacher model using only 8% of its parameters.
• Using GTAN, a student model with 2% of the teacher's parameters achieved a ~2% DSC improvement and 13 ms inference time per 3-D image.

Conclusions:

• The proposed method effectively compresses DL segmentation models using KD and GTAN, maintaining high performance.
• The developed real-time 3-D liver CT segmentation method shows significant potential for liver surgery interventions.
• This approach offers a viable solution for deploying efficient DL models in resource-constrained, time-sensitive medical applications.