NeoRed: A Knowledge-Logic-Alignment Multimodal Large Language Model for Neonatal Respiratory Disease Diagnosis
NeoRed is the first multimodal large language model (MLLM) specifically designed for neonatal respiratory disease diagnosis, addressing the domain gap caused by adult-centric training data in existing MLLMs. The Knowledge-Logic-Alignment (KLA) framework introduces three key components: Knowledge Prior Injection (KPI) for neonatologist-inspired diagnostic priors, Diagnostic Logic Constraint (DLC) for aligning report semantics with diagnostic logic, and Visual Semantic Alignment (VSA) for mapping
Analysis
TL;DR
- NeoRed is the first multimodal large language model (MLLM) specifically designed for neonatal respiratory disease diagnosis, addressing the domain gap caused by adult-centric training data in existing MLLMs.
- The Knowledge-Logic-Alignment (KLA) framework introduces three key components: Knowledge Prior Injection (KPI) for neonatologist-inspired diagnostic priors, Diagnostic Logic Constraint (DLC) for aligning report semantics with diagnostic logic, and Visual Semantic Alignment (VSA) for mapping visual features to imaging conclusions.
- Two new clinical datasets, NeoCXR and NeoCXR-EV, were collected from real-world clinical practice to support training and evaluation of neonatal diagnostic report generation.
- NeoRed achieves a ROUGE-L score of 53.29% and a Clinical Efficacy F1 score of 65.19% on NeoCXR, outperforming existing MLLMs while maintaining competitive performance on adult benchmarks (MIMIC-CXR and IU-Xray).
Why It Matters
This work addresses a critical gap in medical AI by shifting focus from adult to neonatal populations, where disease presentation and imaging characteristics differ substantially. The KLA framework offers a reusable paradigm for domain-adapting MLLMs in specialized medical fields, demonstrating that structured knowledge integration can significantly improve clinical report generation. For AI practitioners working in healthcare, this highlights the importance of domain-specific data collection and knowledge-guided alignment strategies.
Technical Details
- Knowledge Prior Injection (KPI): Incorporates neonatologist-inspired diagnostic priors into multimodal representations, guiding disease-specific attention across clinical context and chest X-ray modalities.
- Diagnostic Logic Constraint (DLC): Constrains the semantics of generated diagnostic reports to align with established multimodal diagnostic reasoning patterns, ensuring clinically coherent outputs.
- Visual Semantic Alignment (VSA): Establishes explicit semantic correspondence between visual features extracted from neonatal chest X-rays and imaging conclusions in diagnostic reports.
- Datasets: NeoCXR and NeoCXR-EV are two real-world clinical datasets collected for neonatal respiratory disease diagnosis; datasets will be available upon application.
- Evaluation: Benchmarked on NeoCXR (ROUGE-L: 53.29%, Clinical Efficacy F1: 65.19%) and adult benchmarks MIMIC-CXR and IU-Xray, demonstrating both neonatal specialization and retained general capability.
Industry Insight
- Domain adaptation in medical MLLMs requires more than fine-tuning; structured knowledge integration (as demonstrated by KPI and DLC) is essential for bridging the adult-to-neonatal domain gap and ensuring clinically valid outputs.
- The dual contribution of new datasets and a specialized model underscores that progress in medical AI increasingly depends on high-quality, domain-specific data collection rather than purely architectural innovation.
- The KLA framework's modular design—separating knowledge injection, logic constraint, and visual alignment—provides a transferable blueprint for adapting MLLMs to other pediatric or specialized medical domains.
Disclaimer: The above content is generated by AI and is for reference only.