advertisement
Editors Selection IGR 12-1
Structure Tests and Anatomy: Machine-learning classifiers for RNFL parameters by Stratus OCT
Comment by Michael Goldbaum on:
25595 Machine learning classifiers for glaucoma diagnosis based on classification of retinal nerve fibre layer thickness parameters measured by Stratus OCT, Bizios D; Heijl A; Hougaard J L et al., Acta Ophthalmologica, 2010; 88: 44-52
Bizios et al. (218) seek to improve the diagnosis of glaucoma by the use of machine-learning classifiers. Why is this important? Healthcare providers rely on automated perimeters and devices that image the optic nerve head and retinal nerve fiber layer to assist in the diagnosis of glaucoma and management of eyes with glaucoma. These perimeters and imaging devices have analysis packages that commonly use standard statistical methods to distinguish eyes with glaucoma from eyes without glaucoma. These statistical classifiers make assumptions about the data that may not be true, and the surfaces that separate the glaucomatous and nonglaucomatous eyes are constrained to specific shapes. These constraints limit the performance of the statistical classifiers. In contrast, machine-learning classifiers adapt to the data when they learn from examples of glaucomatous and nonglaucomatous eyes; the separating surfaces can be formed in the learning process to irregular shapes that maximize the separation, potentially improving diagnosis of glaucoma by reducing the false positives and false negatives. Hence the application of machinelearning classifiers may improve the diagnosis of glaucoma with perimetric and structural imaging devices.
This study applies machine learning classifiers, support vector machine (SVM) with a Gaussian kernel and multilayered perceptron (MLP), to retinal nerve fiber layer thickness (RNFLT) measurements by the Stratus OCT to distinguish between eyes that have glaucoma and eyes that do not have glaucoma. The RNFLT measurements were modified by age and refraction data to improve the usefulness of the thickness measurements. The performance of the machine learning classifiers was enhanced by reducing the number of features used to train and test the classifiers, using a form of feature extraction, local tangent space alignment (LTSA), that preserved neighborhood relationships of similar patterns of RNFLT measurements. The authors found that the SVM and MLP classifiers that they used performed similarly. They used area under the receiver operating characteristic curve (AROC) to measure the performance of the classifiers. AROC of one means perfect ability to separate glaucomatous and nonglaucomatous eyes. With LTSA, they achieved AROC as high as 0.989, which is close to perfect separation. They also found that difference between maximal and minimal thickness of the retinal nerve fiber layer was a good parameter for diagnosing glaucoma.
It is not clear what the authors used to label the eyes in the training and testing sets. This is important, because the indicator chosen to determine which eyes had glaucoma and which eyes were normal for training and testing purposes affects how well the classifiers can be used on new eyes, such as in a clinical setting. Since machinelearning classifiers are potentially an improvement over the methods in the current analysis packages in perimeters and structural imaging devices, a comparison of the MLP and SVM classifiers to the classifiers in current packages would be a useful future study.
Comments
The comment section on the IGR website is restricted to WGA#One members only. Please log-in through your WGA#One account to continue.Log-in through WGA#One
