Graph Flow: Cross-Layer Graph Flow Distillation for Dual Efficient Medical Image Segmentation.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep convolutional neural networks have advanced medical image segmentation.
• High-performance networks require significant computational resources and large annotated datasets, limiting their use in resource-constrained environments.
• Scarcity of annotated medical data hinders the application of complex models.

Purpose of the Study:

• To propose Graph Flow, a comprehensive knowledge distillation framework for network-efficient and annotation-efficient medical image segmentation.
• To address the challenges of high computational costs and limited annotated data in medical image segmentation.

Main Methods:

• Graph Flow Distillation transfers cross-layer knowledge from a teacher network to a compact student network.
• An unsupervised Paraphraser Module purifies teacher knowledge and stabilizes training.
• A unified distillation framework integrates adversarial and vanilla logits distillation for refined predictions.

Main Results:

• Extensive experiments on four diverse medical image datasets (Gastric Cancer, Synapse, BUSI, CVC-ClinicDB) demonstrate competitive performance.
• The framework shows prominent ability across different teacher and student network architectures.
• Effectiveness is validated through a novel semi-supervised paradigm for dual efficient segmentation.

Conclusions:

• Graph Flow offers a dual-efficient solution for medical image segmentation, enhancing both network efficiency and annotation efficiency.
• The proposed framework successfully transfers knowledge from cumbersome models to compact ones, achieving competitive results.
• Graph Flow shows promise for improving medical image segmentation in resource-limited settings and with limited annotations.