Adjustment and validation of convolutional neural network for indirect immunofluorescence interpretation of antinuclear and anti-neutrophil cytoplasmic antibodies: A multicenter study.
Abstract
Objectives
Indirect immunofluorescence (IIF) is the gold standard for autoantibody detection in autoimmune diseases but remains limited by subjective interpretation. We aimed to adjust and validate a standardized convolutional neural network (CNN)-based interpretation system for antinuclear antibody (ANA) and anti-neutrophil cytoplasmic antibody (ANCA) IIF assays across multiple centers.
Methods
The initial CNN algorithm in the EUROPattern (EPA) system was adjusted using 97 ANA and 100 ANCA samples from the lead center. For validation, 935 ANA and 260 ANCA samples from five centers were interpreted by the adjusted CNN-based system and manual interpretation, and inter-method agreement was analyzed.
Results
After adjustment, inter-method agreement for positive/negative classification with manual interpretation improved from 94.8% to 99.0% for ANA and from 84.0% to 96.0% for ANCA; pattern agreement improved from 86.6% to 89.7% and from 84.0% to 96.0%, respectively. In the validation cohort, the adjusted system achieved 98.2% (κ = 0.904) and 95.8% (κ = 0.887) positive/negative agreement for ANA and ANCA, with consistent performance across centers. Overall pattern agreement was 85.6% for ANA and 94.2% for ANCA, with statistically significant discordance only for the nuclear homogeneous pattern in non-mixed samples (P < 0.001); in antibody-specific samples, CNN-based interpretation detected nuclear speckled and cytoplasmic patterns more frequently (P < 0.01). Detection of disease-associated patterns in SLE, PBC, and SS was comparable between methods. Titer agreement within ±1 doubling dilution was 92.8%, 85.1%, and 83.0% for single-pattern ANA, mixed-pattern ANA, and ANCA samples (quadratic weighted κ = 0.775, 0.648, and 0.679, respectively).
Conclusions
A standardized CNN-based system performed comparably to manual interpretation for ANA/ANCA IIF interpretation across multiple centers, showing potential to improve inter-method consistency.