Comparing the Clinical Viability of Automated Fundus Image Segmentation Methods

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Convolutional neural networks (CNNs) are widely used for automatic blood vessel segmentation in fundus images.
• High objective performance metrics do not guarantee clinical applicability of these segmentation masks.

Purpose of the Study:

• To assess the clinical viability of automatically generated blood vessel segmentation masks for diagnosing retinal vascular diseases.
• To compare the clinical quality of segmentation masks generated by different methods.

Main Methods:

• A pilot study involving five experienced ophthalmologists evaluating segmentation masks.
• Ophthalmologists ranked different blood vessel segmentation methods based on clinical quality.
• Correlation analysis between objective performance metrics and subjective clinical evaluation.

Main Results:

• Automatic segmentation masks showed low classification accuracy, limiting their use as a standalone diagnostic tool.
• Ophthalmologists' rankings indicated subjective performance differences among methods with high intra-observer consistency.
• Objective metrics did not correlate with subjective clinical assessments, questioning their utility in selecting clinically robust methods.

Conclusions:

• Current automatic blood vessel segmentation methods are not clinically viable as standalone resources for diagnosing retinal vascular diseases.
• Subjective clinical evaluation is crucial and may not align with standard objective performance metrics.
• Future research should focus on developing methods that correlate objective performance with clinical utility.