Modeling human perception of pterygium fibrovascular redness measurement

Abstract

Pterygium may cause blurring of vision in advanced cases and late treatment may affect the quality of life of a person. The aim of this research is to model a pterygium fibrovascular redness measurement grading scale. The internet enables quick feedback from the experts at the comfort of their home or office. In this study, we demonstrated the use of online form as a tool to get quick feedback from clinicians on clinical grading of pterygium images with various severities. Fifty-one clinicians graded the appearance of thirty images of pterygium fibrovascular redness on a 5-point grading scale with three referent images by an expert. The observers were required to grade each image which was presented in a random order, on a 1 to 3 grading scale. The data collection was analysed by using Statistical Package for the Social Science (SPSS) version 20.0 and Microsoft Excel. The colour space analysis was measured using MATLAB and RAPID MINER Software. The model that we implemented was based on subjective grading by clinicians using descriptive statistics (minimum, 25th percentile, median, 75th percentile, and maximum grade for each of 30 images). The scores were analyzed using quartile analysis and the median was used to construct the benchmark scores for the images. This dataset was tested on assessing human grader and was later trained using artificial neural network to formulate a supervised model using the machine learning algorithm. Intra-class Correlation (ICC) and Bland Altman analyses were performed to assess the performance of human and machine graders. The ICC results for human graders were found to be ranging from 0.57 to 0.89, which indicate poor to excellent agreement with the benchmark scores. The Artificial Neural Network (ANN) exhibits an excellent agreement with an experienced clinician (ICC=0.85), this implies the ANN model was able to mimic the grading of the human expert. This research work has demonstrated the possibility of developing clinical image dataset with its respective grading based on data extracted from an online form. These benchmarked images were shown to be useful in assessing the performance of human and machine learning algorithm. The performance of a newly developed algorithm can also be tested using this dataset in the future.