ScribSD+: Scribble-supervised medical image segmentation based on simultaneous multi-scale knowledge distillation and...

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Deep learning excels at medical image segmentation but requires extensive pixel-level annotations.
• Manual annotation is costly, time-consuming, and requires expertise, limiting deep learning's clinical application.
• Scribble annotations reduce cost but often yield suboptimal segmentation due to limited supervision.

Purpose of the Study:

• To develop a novel framework, ScribSD+, for effective medical image segmentation using cost-efficient scribble annotations.
• To improve deep learning model performance in segmentation tasks with reduced annotation effort.
• To address the challenge of insufficient supervision in scribble-based learning.

Main Methods:

• Proposed ScribSD+ framework utilizing multi-scale knowledge distillation (KD) and class-wise contrastive regularization.
• Employed a student network trained with scribbles and a teacher network using Exponential Moving Average (EMA).
• Implemented multi-scale KD to transfer knowledge from the teacher to the student and contrastive regularization for feature enhancement.

Main Results:

• ScribSD+ significantly enhanced the performance of the student network in medical image segmentation.
• The method outperformed five state-of-the-art scribble-supervised learning approaches.
• Demonstrated effectiveness on ACDC (heart structures) and fetal MRI (placenta, fetal brain) datasets.

Conclusions:

• ScribSD+ offers a powerful solution for learning from limited scribble annotations in medical image segmentation.
• The framework effectively bridges the gap between annotation cost and segmentation accuracy.
• Presents a potential pathway to reduce annotation costs for developing deep learning models in clinical diagnosis.