Knowledge distillation on individual vertebrae segmentation exploiting 3D U-Net | JoVE Visualize

Area of Science:

Medical Imaging
Computer Vision
Artificial Intelligence

Background:

Accurate individual vertebral segmentation on CT scans is crucial for diagnostics and treatment planning in orthopaedics, neurosurgery, and oncology.
Current algorithms face challenges in clinical implementation, particularly regarding integration and computational efficiency.
There is a need for faster, more resource-efficient segmentation methods suitable for clinical settings, including emergency cases.

Purpose of the Study:

To explore the application of knowledge distillation (KD) for training shallower neural networks for efficient vertebral segmentation in CT scans.
To reduce segmentation time and optimize computational and memory resource efficiency.
To assess the performance of KD in improving the accuracy and efficiency of vertebral segmentation models.

Main Methods:

A two-step segmentation approach involving initial spine localization via heatmap prediction and subsequent iterative segmentation.
Implementation of KD by training a teacher network and distilling its knowledge to a shallower student network using soft outputs or logit matching.
Utilized two public datasets with extensive vertebral data and applied data augmentation techniques for robustness.

Main Results:

The teacher network achieved a Dice Similarity Coefficient (DSC) of 88.22% and a Hausdorff Distance (HD) of 7.71 mm.
Knowledge distillation improved the student network's performance, increasing average DSC from 75.78% to 84.70% and decreasing HD from 15.17 mm to 8.08 mm.
The student network demonstrated significant reductions in parameters (75.00%), inference time (36.15%), and CO2 emissions (42.96%) compared to the teacher network.

Conclusions:

This study is the first to apply KD to individual vertebral segmentation in CT scans.
KD enables the development of smaller, efficient neural networks that achieve comparable performance to larger models.
The proposed method offers a feasible solution for improving the efficiency and accessibility of vertebral segmentation in clinical practice.