SCANet: A Unified Semi-Supervised Learning Framework for Vessel Segmentation.

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Accurate subcutaneous vessel segmentation is crucial for clinical venipuncture.
• Deep learning models face challenges in vessel segmentation due to limited annotated data.
• Near-infrared optical apparatus aids in blood vessel localization.

Purpose of the Study:

• To develop a novel semi-supervised learning framework (SCANet) for accurate subcutaneous vessel segmentation.
• To address the scarcity and low-quality of annotated data in medical image segmentation.
• To improve the accuracy of blood vessel localization for clinical venipuncture research.

Main Methods:

• Proposed SCANet framework utilizing a multi-scale recurrent neural network with coarse-to-fine features.
• Incorporated auxiliary branches: a consistency decoder for fine-grained details and an adversarial learning branch for prediction refinement.
• Employed a semi-supervised alternate training strategy with limited labeled and abundant unlabeled data.
• Introduced a new subcutaneous vessel dataset, VESSEL-NIR.

Main Results:

• SCANet achieved accurate segmentation of subcutaneous vessels using a small amount of labeled data.
• The framework demonstrated superiority and generality across various segmentation tasks, including retinal vessels and skin lesions.
• The alternate training strategy effectively leveraged unlabeled data to enhance segmentation performance.

Conclusions:

• SCANet offers a robust solution for subcutaneous vessel segmentation, overcoming data limitations.
• The proposed framework shows significant potential for advancing clinical venipuncture and related medical imaging research.
• The study highlights the effectiveness of semi-supervised learning in medical image analysis.