Application of Model Based Image Interpretation Methods to Diabetic Neuropathy

Abstract

We present two applications of model based computer vision methods to measurement of image features significant in the diagnosis of diabetic neuropathy. The first involves the location of the boundaries of nerve fascicles in light microscope images. The second involves the segmentation of capillary cell regions using electron microscope images. In each case the boundaries required are of arbitrary shape and characterised by local texture or changes in textured regions. The fascicular boundary is located using an Active Contour Model responding to a texture measure based on edge directionality. A start position for the model is automatically generated. The capillary segmentation is performed using a region-based snake responding to a weighted combination of texture measures followed by a local boundary refinement using dynamic programming. These methods show that application of various types of Active Contour Model, accompanied by appropriate starting cues, or followed by local refinements, can locate robustly positioned and intuitively correct boundaries in these images. The aim of the work is the automation of diagnostic measurements currently performed manually. We discuss the implications of automated analysis for procedures in quantitative histology.