Mittuniversitetet

Conventional 3-channel color images have limited information andquality dependency on parametric conditions. Hence, spectral imaging andreproduction is desired in many color applications to record and reproduce thereflectance of objects. Likewise RGB images lack sufficient information tosuccessfully analyze diabetic retinopathy. In this case, spectral imaging may bethe alternative solution. In this article, we propose a new supervised techniqueto detect and classify the abnormal lesions in retinal spectral reflectance imagesaffected by diabetes. The technique employs both stochastic and deterministicspectral similarity measures to match the desired reflectance pattern. At first, itclassifies a pixel as normal or abnormal depending on the probabilistic behaviorof training spectra. The final decision is made evaluating the geometricsimilarity. We assessed several multispectral object detection methodsdeveloped for other applications. They could not proof to be the solution. Theresults were interpreted using receiver operating characteristics (ROC) curvesanalysis.