Abstract
Automation of HEp-2 cell pattern classification would drastically improve the accuracy and throughput of diagnostic services for many auto-immune diseases, but it has proven difficult to reach a sufficient level of precision. Correct diagnosis relies on a subtle assessment of texture type in microscopic images of indirect immunofluorescence (IIF), which so far has eluded reliable replication through automated measurements. We introduce a combination of spectral analysis and multiscale digital filtering to extract the most discriminative variables from the cell images. We also apply multistage classification techniques to make optimal use of the limited labelled data set. Overall error rate of 1.6% is achieved in recognition of 6 different cell patterns, which drops to 0.5% if only positive samples are considered.