A quantitative description of immunofluorescein images using mathematical morphology.

This study investigates whether automatic image shape analysis measurements improved by feedforward neural networks [(FFNN), a software application model] showing the activity of small neural groups can be used as an image analysis tool in the interpretation of skin biopsies evaluated by direct immunofluorescence, by identification of local shape characteristics of patterns. An FFNN software was designed with 4 inputs: the "selected digital characteristics" (SDIC), that is, the proportion of node, link, end point, and branch pixels to total pixels; and 6 outputs: "patterns of direct immunofluorescence images" (PDI), that is, basement membrane linear, basement membrane granular, epidermal intercellular, vascular, negative, and nonspecific background staining; and 10 hidden layers for training process. Microscopic images from anti-IgG, IgA, IgM, C3, and fibrinogen stained sections were collected to a computer; 292 direct immunofluorescence images from 72 patients. The FFNN training set included 192 images from 32 patients. SDIC values and a possibility range for PDI category suggested by an experienced pathologist were transferred to FFNN for training. SDIC values were analyzed statistically according to PDI. After the training phase, for determining the accuracy of automatic self-decision process, 100 images from 40 cases were analyzed by FFNN proposing the possibility of PDI. SDIC values were significantly different for PDI groups (1-way ANOVA; P < 0.001) whereas SDIC values of epidermal intercellular group were different from others. FFNN predicted the correct PDI as the first possibility in 83% and second possibility in 14% of the cases. Direct immunofluorescence SDIC measurements processed by FFNN may help inexperienced pathologists in future.