Multi-scale multi-object semi-supervised consistency learning for ultrasound image segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Manual annotation of ultrasound images is time-consuming and resource-intensive.
• Semi-supervised learning (SSL) leverages unlabeled data to enhance model performance with limited labeled data.
• Existing SSL methods struggle with fusing multi-scale contextual information and handling spatial biases in multi-object segmentation.

Purpose of the Study:

• To introduce a novel semi-supervised learning framework, MSMO, for improved ultrasound image segmentation.
• To address the challenges of multi-scale context fusion and multi-object spatial information bias in SSL.
• To reduce the manual annotation burden in ultrasound image analysis.

Main Methods:

• Developed a consistency learning-based multi-scale multi-object (MSMO) semi-supervised framework.
• Employed a contextual-aware encoder with an attention module for multi-scale feature extraction.
• Utilized a decoder with HConvLSTM for object calibration and recursive multi-object semantic fusion.
• Applied consistency constraints to minimize variations and improve predictions in uncertain regions.

Main Results:

• The proposed MSMO framework significantly outperformed baseline SSL methods on four benchmark datasets.
• Achieved superior performance in both single-object and multi-object ultrasound image segmentation tasks.
• Demonstrated effective fusion of multi-scale contextual and spatial information.

Conclusions:

• MSMO offers a promising solution for automated ultrasound image segmentation, reducing reliance on manual annotation.
• The framework shows significant potential as a clinical tool for medical image analysis.
• The developed MSMO framework enhances the efficiency and accuracy of ultrasound image segmentation.