Related Experiment Video
Updated: Apr 26, 2026
Simultaneous Distinction of Monospecific and Mixed DFS70 Patterns During ANA Screening with a Novel HEp-2 ELITE/DFS70 Knockout Substrate
Published on: January 17, 2018
Automated ICAP AC-30/AC-2 discrimination on HEp-2 indirect immunofluorescence: development and clinical validation of
Changmeng Wu1, Kechi Fang2, Chuan Li3
1Department of Laboratory Medicine, Peking University Shenzhen Hospital, Shenzhen 518000, PR China.
Insights
A new software module accurately distinguishes between AC-2 and AC-30 nuclear patterns in indirect immunofluorescence assays. This tool aids expert-level reporting for autoimmune diagnostics, improving precision and reproducibility.
Area of Science:
- Medical diagnostics
- Immunofluorescence assays
- Autoimmune disease research
Background:
- Distinguishing AC-2 and AC-30 nuclear patterns in HEp-2 indirect immunofluorescence is crucial for ICAP-compliant reporting but challenging.
- The 7th ICAP workshop defined AC-30 as a distinct entity, yet current automated systems lack expert-level discrimination capabilities.
- Accurate pattern recognition is vital as AC-2 interpretation relies on the complete autoantibody profile, not just morphology.
Purpose of the Study:
- To develop and validate a retrieval-based diagnostic software module for expert-level discrimination between AC-2 and AC-30 patterns.
- To address the limitations of current automated systems in differentiating these specific nuclear patterns.
- To enhance the accuracy and standardization of pattern recognition in indirect immunofluorescence assays for ICAP-compliant reporting.
Main Methods:
- A retrieval-based diagnostic software module was created, integrating a MaxViT encoder and a Milvus vector database.
- The module was developed using a retrospective cohort of 557 cases and validated on a prospective clinical cohort of 103 cases.
- Performance was evaluated by comparing six encoder architectures using softmax classification and similarity-based retrieval.
Main Results:
- The optimal module configuration achieved 95.1% overall accuracy, with 100% precision and 90.6% recall for AC-30, preventing false-positive AC-30 assignments.
- Similarity-based retrieval enhanced AC-30 precision and reproducibility across different architectures and runs.
- Each prediction included a retrieved reference image, offering an auditable evidence trail for laboratory quality assurance.
Conclusions:
- A validated retrieval-based diagnostic software module for expert-level AC-2/AC-30 discrimination, aligned with updated ICAP nomenclature, has been presented.
- The system enhances class-specific precision, reproducibility, and interpretability, with a vector database allowing for incremental updates.
- Designed as an assistive tool, the module supports expert interpretation in clinical laboratories, adapting to evolving ICAP standards.
Background:
Accurate discrimination between the AC-2 nuclear dense fine speckled pattern and the AC-30 nuclear fine speckled pattern with mitotic plate on HEp-2 indirect immunofluorescence is required for ICAP-compliant reporting but remains technically challenging. The 7th ICAP workshop established AC-30 as a distinct entity. Current automated systems operate at the competent level and do not address this expert-level distinction. In addition, the clinical interpretation of AC-2 depends on the complete autoantibody profile rather than pattern morphology alone, underscoring the need for precise and standardized pattern recognition.
Methods:
We developed a retrieval-based diagnostic software module for second-stage AC-2 and AC-30 discrimination in specimens with confirmed mitotic plate positivity. The module integrates a MaxViT encoder with a Milvus vector database. It was developed on a retrospective cohort of 557 cases and independently validated on a prospective clinical cohort of 103 cases from a tertiary referral center. Performance was compared across six encoder architectures using both softmax classification and similarity-based retrieval.
Results:
The optimal configuration achieved 95.1% overall accuracy on the internal hold-out validation set, with 100% precision and 90.6% recall for AC-30, resulting in no false-positive AC-30 assignments. Similarity-based retrieval consistently improved AC-30 precision across all evaluated architectures and demonstrated enhanced reproducibility across repeated runs. Each prediction was accompanied by a retrieved reference image, providing an interpretable and auditable evidence trail for laboratory quality assurance.
Conclusion:
This study presents a validated retrieval-based diagnostic software module for expert-level AC-2/AC-30 discrimination under updated ICAP nomenclature. The system improves class-specific precision while maintaining reproducibility and interpretability. Its vector database design enables incremental updates without retraining, supporting practical implementation in clinical laboratories as ICAP standards evolve. The module is designed as an assistive tool to support expert interpretation rather than replace clinical decision making.
