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

Optical coherence tomography angiography (OCTA) offers dye-free, three-dimensional views of skin microvasculature, yet progress in developing reliable and quantitative solutions for vessel architecture analysis is slowed by heterogeneous preprocessing practices, scarce annotated data, and limited evaluation metrics.

We assess how typical OCTA preprocessing steps influence the accuracy of deep learning vessel segmentation, and we identify network designs and metrics best suited to OCTA dermatological data.

Experiments use the open DERMA-OCTA dataset containing 330 volumes from different skin conditions; each volume is additionally provided in five progressively pre-processed versions: original, Bscan normalization, projection artifact attenuation, contrast enhancement, and vesselness filtering. Segmentation is performed with representative 2D and 3D deep learning approaches. Besides standard segmentation metrics, evaluation includes the connectivity-area-length index, which proved particularly effective for assessing dermatological vessel segmentation.

The analysis shows that Bscan normalization, projection artifact attenuation, and contrast enhancement incrementally improve segmentation accuracy, whereas vesselness enhancement can impair segmentation performance. Among the tested architectures, 2D models achieved the highest segmentation performance, although 3D approaches proved more effective for deeper tissue layers. Testing across different pathologies revealed challenges in model generalization to varied vascular patterns.

Combining 2D and 3D models and using topology-aware indices provide a full, clinically relevant evaluation of algorithm performance.