Low-Quality Sensor Data-Based Semi-Supervised Learning for Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Traditional medical image sensors require extensive labeled data, which is costly and time-consuming.
• Limited computational power hinders effective model training and inference with large datasets.
• Poor generalization across different data modalities is a significant challenge for existing sensors.

Purpose of the Study:

• To develop a novel framework, LSDSL, for semi-supervised learning in medical image segmentation using low-quality sensor data.
• To address the limitations of traditional medical image sensors in data acquisition, computational power, and generalization.

Main Methods:

• LSDSL employs semi-supervised learning on low-quality sensor data for medical image segmentation.
• The hard region exploration (HRE) module enhances model understanding of low-quality pixels in challenging regions during supervised learning.
• A pseudo-label sharing (PS) module facilitates knowledge transfer from high-quality to low-quality pixels across networks in unsupervised learning.

Main Results:

• The LSDSL framework demonstrates superior performance compared to existing semi-supervised methods.
• Achieved enhanced segmentation accuracy and faster inference speeds on CT and MRI datasets.
• Successfully adapted to datasets from different modalities, showcasing improved generalization.

Conclusions:

• LSDSL offers an effective solution for medical image segmentation challenges, particularly with low-quality data.
• The framework improves model performance and efficiency in medical imaging applications.
• LSDSL shows promise for broader application in medical image analysis across various modalities.