Background and Aims

Image analysis is used for evaluation of immunofluorescence assays (IFA), e.g. detection of antinuclear antibodies (ANA) on HEp-2 cells. Automated IFA interpretation systems can capture these images, calculate relative fluorescence intensities and ascertain the pattern. Machine learning algorithms, partly referred to as artificial intelligence (AI), are considered the state-of-the-art in image classification. Especially deep learning algorithms based on convolutional neural networks (CNN) are regarded particularly powerful. The aim of the study was to create a complete AI-supported model for the classification of ANA patterns according to ICAP nomenclature.

Methods

CNNs were established and trained for major ANA sub-classification tasks: a) metaphase recognition in DAPI, b) metaphase pattern, c) interphase pattern and d) cytoplasmic pattern. Images were obtained by an automated IFA interpretation system (AKLIDES, Medipan, Germany).

Sets of pre-classified images were created for each task: a) training set for supervised learning b) validation set and c) test set for evaluation of prediction accuracy.

Results

Trained CNNs showed true classification rates above 95% for all models. The CNN were able to learn important features for major ANA sub-classification tasks. Merged results provided a classification model for ICAP nomenclature tree.

Conclusions

Machine learning algorithms add value to enhance the accuracy of image classification in automated IFA interpretation. In addition to classical pattern-recognition methods, these algorithms can provide extended information for the improvement of classification algorithms. Though our approach showed efficient learning, it required an adapted image capturing strategy.

This work was supported by BMBF - PRÆMED.BIO research grant (03WKDB).