An automated glaucoma risk indicator using digital color fundus photos proved to be accurate in the detection of glaucoma disease in a recent study. This novel technique is not intended to replace other state-of-the-art diagnostic techniques currently used . Ophthalmologists, however, can receive a confirmation from this automated diagnosing system that the diagnosis they are making in regards to glaucoma is the right one.
Glaucoma remains one of the leading causes of blindness and is characterized by the progressive loss of retinal nerve fibers in the parapapillary region. The loss of these nerve fibers as the disease progresses is permanent, underscoring the importance of detecting the disease early. State-of-the-art retina and retinal nerve fiber imaging devices (such as the Heidelberg Retinal Tomograph 3 by Heidelberg Engineering and newer high-resolution optical coherence tomography [OCT]) are widely used to assist ophthalmologists in the early diagnosis of glaucoma disease, in addition to the manual diagnoses made using an ophthalmoscope.
This new automated glaucoma detection system is designed to serve as yet another useful tool in the accurate diagnosis of early glaucoma, ultimately offering a more favorable prognosis to patients with early or borderline glaucoma.
The study included the papilla-centered digital color fundus images of 100 glaucoma patients (FDT test time 67.25±33.4 s) and 100 patients with normal eyes using a non-mydriatic (FOV 22.5°) digital camera (Kowa). A pre-processing step of the system eliminated certain disease-independent variations from the input images such as illumination inhomogeneities, papilla size differences, and vessel structures. To characterize glaucomatous changes, generic feature types, including pixel intensities, frequency coefficients, histogram parameters, Gabor textures and spline coefficients, were extracted.
Results showed that the classification of compressed raw pixel intensities had a success rate of 83%, with a specificity of 0.72 and a sensitivity of 0.94, in detecting glaucomatous cases. This success rate in reaching an accurate diagnosis of glaucoma increased to 86% when using spline coefficients with a specificity of 0.78 and a sensitivity of 0.94; the combination of both features slightly increased the specificity to 0.82.
"The system we used for determining the risk index for glaucoma is a fully automated system," Meier said. "This new, data-driven approach achieves an accurate detection of glaucomatous retina fundus images comparable to ophthalmologists."
Digital color fundus photos are one of the most important modalities used in glaucoma detection to get a first look at the retina, and they are easy to acquire compared with the other imaging techniques currently used in ophthalmology such as laser scanning tomography. According to Meier, the cameras used are much less expensive than laser scanning tomography or high-resolution OCT examinations and, therefore, are much more widely found and used in private practice and in clinics.
"The pictures of the retina of patients along with their respective diagnoses were used to train the classifier in the system, and, ultimately, a training set of glaucoma images was stored in the system's database," Meier said. "The accuracy of this training set was then evaluated by taking another set of images in our study to see how our risk index for glaucoma corresponds to the diagnoses made by the ophthalmologist."
Most experts would agree that glaucoma is difficult to characterize in detail, even for experienced ophthalmologists. Other classification methods use segmentation measurements in which the algorithm tries to identify certain physiologic structures and outline them according to measurements (such as size), and these numbers then are taken as a basis of classification. According to Meier, one major drawback of segmentation-based techniques is that small errors in segmentation may lead to a significant change in the measurements and thus the estimation and diagnosis. The novel automated technique takes a more data-driven classification approach and uses an appearance-based method for the detection of glaucoma.
"This approach is not intended to replace the expertise of the ophthalmologist nor replace the invaluable diagnostic capabilities of other imaging techniques currently used and high-resolution OCT but instead offers the ophthalmologist an additional, automated marker for a more secure diagnosis of glaucoma," Meier said.
Morphologic changes occur in the retina of patients before the patient ever suspects or presents with symptoms of the disease. To a certain extent, the loss of retinal nerve fibers can be compensated by the eye and by the brain. The changes there already can be seen in the images of the retina before these changes become apparent to the patient.
The tomographer is designed to image the topography of the retina, which gives a clinician the height information of the physiologic structures and the papilla. The high-resolution OCT also gives a clinician this height information as well as vital information concerning the status of the retinal nerve fiber and its thickness, the latter being crucial in determining the prognosis of glaucoma.
Meier said that not every ophthalmologist may be a "glaucoma expert," and this automated diagnostic approach may help them significantly in reaching the precise diagnosis in their patients. This approach also can be particularly useful for younger ophthalmologists who lack the clinical experience concerning glaucoma disease, again, helping them reach the precise diagnosis and can serve parallel as a learning device for glaucoma disease recognition, he said.