Impact of image preprocessing on dermatological OCTA vessel segmentation: a DERMA-OCTA study

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical coherence tomography angiography (OCTA) provides non-invasive, 3D visualization of skin microvasculature.
• Challenges in OCTA analysis include varied preprocessing, limited data, and inadequate evaluation metrics for vessel architecture.

Purpose of the Study:

• To evaluate the impact of preprocessing techniques on deep learning-based vessel segmentation in OCTA.
• To identify optimal deep learning architectures and evaluation metrics for dermatological OCTA data.

Main Methods:

• Utilized the DERMA-OCTA dataset with 330 volumes and five preprocessing variations.
• Applied 2D and 3D deep learning models for vessel segmentation.
• Evaluated performance using standard metrics and the connectivity-area-length index.

Main Results:

• Bscan normalization, artifact attenuation, and contrast enhancement improved segmentation accuracy.
• Vesselness filtering negatively impacted performance.
• 2D models excelled overall, while 3D models were better for deeper tissues.
• Models faced generalization challenges across diverse pathologies.

Conclusions:

• Combined 2D/3D models offer comprehensive analysis.
• Topology-aware indices provide clinically relevant performance evaluation for OCTA vessel segmentation.