Background The gold standard for ANA screening is indirect immunofluorescence (IIF) on HEp-2 cells. Automation of IIF image reading can provide a reliable basis for cost-effective serological diagnostics. Recentely there has been an increase demand of computer-aided diagnosis (CAD) tools to support clinicians in the field of IIF. The novel digital imaging reading approach can help us to overcome the reader subjectivity, allowing a permanent archival record of the results and CAD use. We described a CAD system (SLIM-system) for ANAIIF images reading (1).
Objectives To assess inter-observer variability using digital images reading approach instead of manual visual approach.
To compare accuracy of expert readers of three different laboratories to automated (CAD) classification.
Methods Three laboratories (named A, B, C) expert in IIF field were involved in the multicenter project which recruited 305 consecutive serum samples submitted to laboratory for routine ANA testing. Every image has been classified from experienced physicians.
We acquired 915 images using HEp-2 cell slides (BION-DASIT) at 1:80 screening dilution as recommended by the current guidelines. Each expert took 3 different images per sample with an acquisition unit consisting of the fluorescence microscope coupled with a LED light source and a digital camera. The images with a resolution of 1388x1038pixels, color depth 24 bits were stored in bitmap format. We used 40-fold magnification.
Digital images have been reclassified from two blinded experts into three intensity classes: positive, negative, weak positive. Thus we compared obtained data with CAD classification.
Data have been analyzed in a pairwise manner.
Agreement have been performed by Cohen's kappa test
Results Fig1 shows results. We obtained 99.6%, 69.48% and 29.55% of sensitivity on positive, negative and weak positive images of respectively comparing A and B; 92.74%, 60.49% 49.24% comparing C and A. We evaluated accuracy and specificity of visual reading: 71.65% and 69.48% respectively comparing group B-A; 67.39% and 60.49% comparing group C-A. A-B, B-C and A-C agreement were 0.551 (±0.0237), 0.69 (±0.02) and 0.509 (±0.023) respectively The SLIM-system image classification showed a sensitivity of 92.03%, 94.94%, and 43.41% on positive, negative and weak positive, respectively Comparing SLIM-system results to average experts' classification we obtained: accuracy 86.87%vs69.52%, specificity 94.94%vs64.98%.
Conclusions In literature there are no studies on IIF interobserver reading variability by digital images reading approach. We have compared interobserver variability among 3 laboratories experienced in ANA field. The operating principle of new automated systems is acquiring, storing, and analyzing of digital images of stained IIF slides, and is needed to standardize image acquisition. SLIM-system data analysis showed a higher sensitivity on negative and weak positive images classification compared to the manual visual approach. The new CAD systems can eliminate the need for a darkroom and separate image acquisition from image interpretation, besides have the potential to improve the quality and utility of the ANA HEp-2 assay.
Rigon A et al Autoimmun Rev. 2011 Aug;10(10):647-52.
Disclosure of Interest None declared