Quantitative comparison of segmentation methods for in-body images

Area of Science:

• Medical Imaging
• Computational Pathology
• Gastroenterology

Background:

• Accurate segmentation of endoscopic images is crucial for clinical analysis.
• Identifying critical areas in medical images requires precise delineation.
• Manual segmentation by clinicians is the current standard but is time-consuming.

Purpose of the Study:

• To numerically compare mean shift and normalized cuts segmentation algorithms.
• To evaluate how well these algorithms approximate expert gastroenterologist annotations.
• To determine the optimal algorithm for segmenting vital-stained magnification endoscopy images.

Main Methods:

• Quantitative comparison of mean shift and normalized cuts algorithms.
• Application of algorithms to a set of endoscopic images.
• Comparison of algorithm results against manual segmentations by two specialist clinicians.

Main Results:

• Normalized cuts performed better when considering only the most relevant single segmented patch.
• Mean shift demonstrated superior performance when the annotated area could be represented by multiple patches.
• The need for automatic methods to determine the kernel's bandwidth for mean shift was highlighted.

Conclusions:

• The choice of segmentation algorithm depends on the desired representation of annotated areas.
• Mean shift offers advantages for complex annotations requiring multiple patches.
• Further research into automated parameter selection for mean shift is warranted for clinical application.