Statistical analysis of retinal tomographic pseudo images for diagnostic purpose

The majority of visual impairments are related to retinal disorders. Therefore, techniques useful for early detection of retinal anomalies are of great interest for diagnosis and treatment of related pathologies. Optical Coherence Tomography (OCT) is a non-invasive technique that is becoming increasingly important in the analysis of human retina. This technique produces sections of a three dimensional reflectance map, which are pseudo-images of the retinal structure. In these pseudo-images, early pathologies may produce local modifications of reflectance, which visually, appear as blots. Consequently, automatic procedures for detection of such blots may be of help in the diagnosis of many retinal diseases. Due to the complexity of the measurement technique and of the retinal structure, the pseudo-images produced by OCT can be modeled as 2D random signals: the above pathological blots correspond to sudden changes in the distribution of such OCT signals. In this work, we performed a statistical analysis of retinal OCT data. We found that the so-called stretched-exponential distribution, a two-parameter probability density function, provides a model that describes the data in a satisfactory way. In addition, we analyzed spatial-reflectance interdependence by studying the autocorrelation function; this analysis showed a significant correlation between contiguous pixels about 5 μm apart. We then developed a procedure to obtain the maximum-likelihood estimate of distribution parameters along the retinal pseudo-image. With this procedure, we analyzed their spatial dependence both along and across retinal layers. We observed that along suitable thin layers, the two parameters were statistically constant. By exploiting this result, a statistical procedure was developed for automatic detection of parameter variations due to retinal pathological blots. Some first results are reported for detection of diabetic retinopathy.