Experiences of Digitisation in HEp-2 Pattern Recognition – from Human to Machine

A Brief Overview

Human epithelial (HEp-2) cells are a vital tool in the diagnosis of autoimmune diseases. In recent years, there have been advancements to improve standardisation and quality of autoimmune diagnosis using indirect immunofluorescence assay (IIFA) on HEp-2 cells.  

Many laboratories have switched over to using automated systems, that not only perform the assays but also help interpret the results using pattern recognition algorithms. The introduction of automation into busy labs has reduced the processing time [1] of patient samples, as well as the risks of human error during the handling of samples. Automated processors and microscopes were also developed to improve standardisation reducing intra- and inter-laboratory variation.

However, not all have accepted this new technology easily, raising concerns that it takes away the skills and knowledge from the experienced lab and clinical professionals. Others see the benefits of increasing throughput and promoting standardisation. Yet, some of these automated systems still require a trained professional to manually check all images collected by the digital microscope and are still therefore not fully automated.

So what does the future hold for HEp-2 pattern recognition and diagnosis? Will there be a complete switch to automation?

A History of IIFA using HEp-2 cells

HEp-2 cells are used extensively by diagnostic laboratories worldwide for the diagnosis of a wide range of autoimmune diseases. When used in IIFA, these cells allow for the visualisation of anti-nuclear antibodies (ANA) as well as cytoplasmic antibodies in patient serum. Specific ANA patterns and titres are associated with various systematic autoimmune rheumatic diseases (SARD) [2] such as Sjögren’s syndrome (SjS), systemic sclerosis (SSc), mixed connective tissue disease (MCTD), and systemic lupus erythematosus (SLE).

IIFA on HEp-2 cells has been used for many years but is still considered the gold standard [3] for autoantibody diagnosis in SARD. As well as diagnosing the presence or absence of ANA antibodies, HEp-2 IIFA also helps determine the end-point titres of samples and therefore helps their clinical relevance.

Reading HEp-2 slides down a microscope involves a lot of skill that takes years of experience to master due to the range of patterns that can be detected. It can be a slow process with a high rate of subjectivity, which has led to a lack of standardisation in the past.  Correct diagnosis is vital, not only for correct follow-up testing but also to allow for the best possible care and treatment for the patient.

Introducing Standardisation and Harmonization of ANA patterns on HEp-2 Cells

More recently, the International Consensus on ANA Patterns or ICAP [4] was introduced to apply consensus to the variations of the vast number of morphological patterns seen in HEp-2 IIFA. It aims to apply standardisation [5] across the difficult area of pattern recognition and diagnosis. Currently, there are 29 [6] recognised HEp-2 IIFA patterns. Each distinct pattern has a nomenclature to allow standardisation. The titre and the pattern of the result are important for patient diagnosis.

The ICAP nomenclature has also been recently applied to some automated systems, in an attempt to ease the manual issue of HEp-2 IIFA subjectivity in pattern recognition.

The Introduction of Sample Processing Platforms and Digital Microscopy

So when did slide processors and readers become part of autoimmune diagnostics? Automated systems for slide processing came first, having been around for many years. However, interpreting the slides was left to the trained professional. It is only in the last decade that digital microscopes attached to slide processors have become more widely available. Due to the complex nature of HEp-2 patterns and their clinical relevance, this area of automation has been somewhat slower to develop than in other diagnostic areas.

Initially, HEp-2 IIFA was labour intensive. Automation replaced this process, freeing up time for laboratories. However, the processed slides would still need to be read manually by a trained individual under the microscope, due to the complexities in reading the slides. Furthermore, new high-throughput assays became available such as ELISAs (enzyme-linked immunosorbent assays), line immunoassays (LIAs), and multiplexed bead assays (MBAs). However, variations [7] in conjugates, antigen source, purity and concentration, reference material, and other components used by various manufacturers led to differences in results between immunoassays for the same patient.

The introduction of automated digital microscopes has come with mixed reviews. Some clinicians and scientists were happy with the switch from microscope to digital images, others, however, were unhappy with the change. How can a computer navigate correctly over an image? How can it adjust its focus to obtain a 3D image which is often helpful in aiding interpretation? Furthermore, the clinician does not only look at one section of the well, it is important to look at the whole area before making a decision. It takes trust to hand over these skills to a computer.

Newer advancements include the release of automated computer-assisted pattern recognition systems. These systems obtain digital images of processed HEp-2 slides then store and analyse them. There are currently many automated systems available [7] for HEp-2 IIFA. They vary on factors such as processing time, counterstains, the number of patterns that can be identified by the software, titre prediction, and others. It is not uncommon for the systems to have an in-built pattern image library which not only helps when assigning patterns to samples but can also be used as training aids. At present, however, all images must be checked and signed off by trained professionals.

Human Vs Machine

Studies have been carried out to determine how automated IFA system image analysis compares with manual procedures [8]. Some issues have been found with even basic pattern recognition by the digital readers, but the overall performance of correctly diagnosed positive and negative samples is comparable. With less common patterns, however, or those that are difficult to interpret, the digital reader will sometimes struggle. Some false negatives have been detected with positive cytoplasmic samples, nuclear dots, or nucleolar patterns because these emit fluorescence in small areas, making the overall fluorescence appear low. However, as the operator must digitally sign off on each image before the results can be sent off, incorrectly reported results are unlikely. Operators can also tweak the settings to their preference. 

Unfortunately, there are still differences between manufacturers [9] in terms of conjugates and HEp-2 substrates used. This may affect standardisation across laboratories. Computer-aided diagnostic systems (CAD) [8, 9] have been developed to overcome the lack of standardisation for HEp-2 pattern recognition and titre. Manually, it is subjective, using CAD systems it is hoped that there will be more harmonisation in the field of autoimmunity. Further reducing intra- and inter-laboratory variability.

What is the Future for Hep-2 IIFA?

The manual review of HEp-2 images post-processing may soon be replaced by intelligent learning technology (unsupervised learning paradigm)[10]. This technology will correctly diagnose samples and submit results, saving time manually labelling patient HEp-2 images. An enhanced learning scheme has been proposed which includes a completely unsupervised paradigm. The use of a deep convolutional auto-encoder (DCAE) [10] may be used to analyse images and extract and discriminate pattern information. It will include a clustering layer that can discriminate during the feature learning process, the hidden pictures formed by the DCAE.

If new technology further enhances standardisation across laboratories for HEp-2 IIFA and improves patient outcome, then it would be a welcomed addition. However, will this new technology lead to a decline in human expertise in HEp-2 IFA pattern recognition? For those skilled in the area, the change may be unwelcome. This, along with the difficulties that manufacturers have faced in programming software to interpret HEp-2 patterns could be the reason that full automation of autoimmune diagnosis in IIFA has trailed behind other areas of diagnostics.


[1] Voigt, J., Krause, C., et al.  (2012). Automated indirect immunofluorescence evaluation of antinuclear autoantibodies on HEp-2 cells. Clinical & developmental immunology. ID: 651058. DOI: 10.1155/2012/651058

[2] Mahler, M., Meroni, P-L., Bossuyt, X., et al. (2014) Current concepts and future directions for the assessment of autoantibodies to cellular antigens referred to as anti-nuclear antibodies. J Immunol Res 2014:1–18.  DOI: 10.1155/2014/315179

[3] Damoiseaux, J., et al. (2019) Clinical relevance of HEp-2 indirect immunofluorescent patterns: The International Consensus on ANA patterns (ICAP) perspective. Annals of the Rheumatic Diseases. (78) 879-889.  DOI:10.1136/annrheumdis-2018-214436

[4] ICAP, International Concensus on ANA patterns. (2021). https://anapatterns.org/index.php

[5] Chan Edward, K. L., Damoiseaux, J., et al. (2015). Report of the First International Consensus on Standardized Nomenclature of Antinuclear Antibody HEp-2 Cell Patterns 2014–2015. Frontiers in Immunology. Aug 20;6:412.  DOI: 10.3389/fimmu.2015.00412  

[6] ICAP, Nomenclature and Classification Tree/ (2021). https://anapatterns.org/trees-full.php

[7] Tebo, A,E. (2017). Recent Approaches To Optimize Laboratory Assessment of Antinuclear Antibodies. Clin Vaccine Immunol. 5;24(12):e00270-17. DOI: 10.1128/CVI.00270-17.

[8] Infantino, M., et al. (2017). The burden of the variability introduced by the HEp-2 assay kit and the CAD system in ANA indirect immunofluorescence test. Immunol Res. Feb;65(1):345-354. DOI:10.1007/s12026-016-8845-3.

[9] Cinquanta, L., Bizzaro, N., and Pesce, G. (2021). Standardization and Quality Assessment Under the Perspective of Automated Computer-Assisted HEp-2 Immunofluorescence Assay Systems. Frontiers in Immunology. (12) page 218. DOI: 10.3389/fimmu.2021.638863    

[10] Vununu, C., Lee, S-H., Kwon, K-R.  (2020). A Strictly Unsupervised Deep Learning Method for HEp-2 Cell Image Classification. Sensors. 20(9):2717. DOI: 10.3390/s20092717.

Share This Post