Workflow evaluation of a commercial deep learning-based image analysis tool for the quantification of angiogenesis in vivo

Angiogenesis plays a crucial role in tumor development as well as in wound healing and integration of biomaterials. However, quantification of complex vascular networks across different in vivo models is time-consuming, labor-intensive, and prone to human error. Deep learning-based analysis (DLBA) appears to be a more efficient and reproducible method for vascular quantification . Yet, studies directly comparing its performance with human assessment remain limited. IKOSA CAM is a commercially available tool built for analysis of vascular networks. Here we evaluated IKOSA CAM against manual vessel tracing across three different angiogenesis models. Images obtained from three different in vivo angiogenesis models, the chorioallantoic membrane (CAM) assay, the rodent retina, and the dorsal skinfold chamber (DSC), were analyzed. The vascular network was analyzed using both manual tracing and the commercial deep learning-based image analysis tool IKOSA CAM. In addition, tracing results were reevaluated by human raters and the total workflow duration was assessed. Deep learning-based image analysis proved to be a time-efficient method for tracing and quantifying vascular networks. The level of agreement with manual human tracing varied among image types. The highest agreement between the two methods was observed for the CAM assay. The greatest differences occurred in the tracing of DSC branching points, as indicated by a downward trend in the Bland-Altman plots. Repeated analyses with the DLBA produced identical results for all tested angiogenesis models, confirming excellent reproducibility. Standardized vascular quantification remains a major challenge in experimental angiogenesis research. In this workflow validation study, the evaluated DLBA provided reproducible and time-efficient vessel quantification, particularly for the CAM assay, while showing limitations in the analysis of retinal and DSC images. Overall, DLBA represents a promising approach for standardized vascular analysis and may facilitate reproducible, high-throughput image quantification. Nevertheless, further validation using larger and more heterogeneous datasets is warranted.