G-MBRMD: Lightweight liver segmentation model based on guided teaching with multi-head boundary reconstruction...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate liver segmentation is crucial for liver cancer analysis.
• Current deep learning models are computationally expensive, hindering clinical use.
• There is a need for efficient and lightweight liver segmentation solutions.

Purpose of the Study:

• To propose a real-time, lightweight liver segmentation model named G-MBRMD.
• To enhance segmentation performance without increasing computational complexity.
• To enable widespread practical application of liver segmentation technology.

Main Methods:

• Employed knowledge distillation with a Transformer-based teacher and a convolution-based student model.
• Introduced multi-head mapping and boundary reconstruction strategies.
• Guided the student model to learn global boundary processing from the teacher.

Main Results:

• Achieved an average Dice coefficient of 90.14±16.78% on the LITS dataset.
• Model size is only 0.6 MB with an inference speed of 0.095s on a standard CPU.
• Improved the baseline student model's Dice coefficient by 1.64% without added computational cost.

Conclusions:

• The G-MBRMD model successfully unifies segmentation precision and model lightness.
• The method significantly enhances the potential for widespread application in clinical settings.
• This lightweight model addresses the computational limitations of current deep learning segmentation techniques.