Goblet cells segmentation from confocal laser endomicroscopy with an improved U-Net

Area of Science:

• Gastroenterology and Medical Imaging
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Gastric intestinal metaplasia (GIM) is a key precursor to intestinal-type stomach cancer.
• Accurate assessment of GIM relies on goblet cell (GC) segmentation from confocal laser endomicroscopy (CLE) images.
• Manual GC segmentation is labor-intensive, prone to errors, and struggles with CLE image quality.

Purpose of the Study:

• To develop an automated and accurate method for GC segmentation in CLE images.
• To improve the assessment of GIM for early stomach cancer detection.
• To overcome limitations of manual segmentation and existing automated approaches.

Main Methods:

• Collected and manually annotated 343 CLE images from 62 patients.
• Developed an improved U-Net model incorporating a pixel gradient attention mechanism.
• The model focuses on color gradient information for enhanced feature map guidance.

Main Results:

• The proposed GCSCLE method achieved high segmentation performance.
• Achieved an Intersection over Union (IOU) of 87.95% and a Dice coefficient of 86.64%.
• Demonstrated comparable performance to manual segmentation in clinical settings.

Conclusions:

• The GCSCLE method offers a reliable and efficient alternative to manual GC segmentation.
• This automated approach can significantly improve segmentation accuracy, saving time and costs.
• Facilitates better assessment of GIM and aids in early stomach cancer diagnosis.