Purpose:
To develop a computer algorithm to estimate vessel tortuosity in retinal images acquired with wide-field fundus cameras in Retinopathy of Prematurity (ROP) infants. It will allow clinicians to recover an objective quantitative estimation of tortuosity.
Methods:
130° retinal images were acquired with RetCam (Clarity Medical Systems, Pleasanton, CA) in 20 normal and ROP (pre-plus and plus) infants. Given the difficulty for humans to provide a quantitative assessment of tortuosity, 3 clinical graders and 3 ROP image experts independently ordered the images by increasing tortuosity. In order to focus on tortuosity estimation alone and avoid problems from possible errors in computerized vessel tracing, a manual tracing of visible vessels was carried out by one of the authors and used for computer tortuosity estimation.Various indices of tortuosity, based on different geometrical features, were assessed, both single and combined. The selection of the best weights to combine indices was done using a "leave-one-out" technique. One image out of 20 was chosen, the remaining 19 were used to find the optimal weights of the combination under exam and the resulting combination was assessed on the left-out image. The whole procedure was repeated 20 times, each time leaving-out a different image, eventually yielding the combined index value for all the 20 images.The criterion to select the best (single or combined) index was the correlation of its ordering with ground truth (clinical graders’ average ordering).
Results:
The table reports the Spearman Correlation Coefficient for each pair of ordering. A remarkable improvement is shown when using combined over single index, with combined index performing at clinical graders’ level.
Conclusions:
The proposed algorithm was shown to behave in this dataset at a level comparable to the best experts' (clinical graders). When paired with an automatic vessel tracing technique (currently in late stage of development), it will provide a completely automated tool for the reliable quantitative estimation of vascular tortuosity in ROP images.
Keywords: imaging/image analysis: clinical • retinopathy of prematurity • image processing