Abstract
Abstract: :
Purpose: RetCam images from patients with retinopathy of prematurity (ROP) vary largely on their radiometric parameters. Furthermore viewing conditions differ in remote settings, because of monitor quality, calibration profiles and ambient light. Since there are no current standards, we sought to automate and objectively optimize the image characteristics. Methods: Images from routine screening tests of patients with ROP were selected to represent a spectrum of radiometric variations. Images were viewed on high resolution color LCD monitors. To assure standardization of the viewing parameters, the monitors were calibrated with a hardware–based calibration system and no ambient light present. Images were stratified in separate categories and median radiometric conditions were identified and assigned to a group of images as optimal. Manual adjustment of brightness and contrast to each image was attempted and the settings recorded to a database. A measurement software algorithm was designed to automatically determine the image intensity levels at certain image locations using matrixes. The measurements were performed to the optimal as well as to the non–optimal images and the differences calculated before and after the adjustment. Based on the calculated values and the recorded manual adjustments a second algorithm using fast programming languages was developed to compensate for the difference, if the image is a part of non–optimal category. One hundred randomly selected images from a routine screening for ROP were tested for robustness and accuracy. Results: The algorithms showed reliability in the performance tests on low and high–end computer workstations. Because of the VGA image resolution and matrix enabled measurement, the evaluation and adjustment cycle takes between 15–30 milliseconds. The automatic adjustment yielded 100 % reproducibility, when the same images were presented to the software. Two graders established their preference profile initially and let the software perform automatic adjustments on 100 images. The automatic adjustment agreed with their manual in 98% and 95% of the time. They also stated the quality of the automated adjustment as better, the same or worse than the unadjusted image. None of the images fell in the worse category after the automated adjustment and the distribution of the better and the same category was 82 % and 18 % respectively. Conclusions: Robust reproducible algorithms were developed for accurate automatic image adjustment of RetCam images. After establishment of an user defined profile, the software performs unassisted automatic adjustment.
Keywords: image processing • clinical (human) or epidemiologic studies: systems/equipment/techniques