Deep Learning Fuzzy Inference: An Interpretable Model for Detecting Indirect Immunofluorescence Patterns Associated with Nasopharyngeal Cancer

Abstract

The detection of serum Epstein-Barr virus antibodies by immunofluorescence assay (IFA) is considered the gold standard screening test for nasopharyngeal cancer (NPC) in high-risk populations. Given the high survival rate after early detection in asymptomatic patients, compared to the poor prognosis in patients with late-stage NPC, screening using IFA has tremendous potential for saving lives in the general population. However, IFA requires visual interpretation of cellular staining patterns by trained pathology staff, making it labor intensive and hence nonscalable. In this study, an automated fuzzy inference (FI) system achieved high agreement with a human IFA expert in identifying cellular patterns associated with NPC (κ = 0.82). The integration of a deep learning module into FI further improved the performance of FI (κ = 0.90) and reduced the number of uncertain cases that required manual evaluation. The performance of the resulting hybrid model, termed deep learning FI (DeLFI), was then evaluated with a separate testing set of clinical samples. In this clinical validation, DeLFI outperformed human evaluation on the area under the curve (0.926 versus 0.821) and closely matched human performance on Youden J index (0.81 versus 0.80). Data from this study indicate that the combination of deep learning with FI in DeLFI has the potential to improve the scalability and accuracy of NPC detection.