4DfCF AI model delivers transparent, efficient brain-disorder screening.

A research team at Sungkyunkwan University’s InfoLab, led by Professor Tamer Abuhmed, has developed a novel vision transformer architecture, the 4D fMRI CrossFormer (4DfCF), to enhance the transparency and computational efficiency of AI-driven brain disorder screening. Published in the IEEE Journal of Biomedical and Health Informatics, the model addresses a persistent challenge in neuroimaging: functional magnetic resonance imaging data captures dynamic brain activity across both space and time, rendering it difficult to analyze with conventional deep learning frameworks. Unlike standard models that process static brain scans, 4DfCF analyzes continuous three-dimensional image sequences to identify spatial and temporal activity patterns simultaneously. The architecture operates across multiple scales, capturing both localized neural interactions and long-range connections throughout the brain. This dual-focus approach enables the system to detect subtle, disease-specific signatures that traditional methods frequently overlook. A defining characteristic of 4DfCF is its integration of explainable AI techniques. By generating interpretability maps, the system highlights the specific brain regions that contributed most strongly to its diagnostic predictions. Rather than delivering opaque binary classifications, the framework provides clinicians with visual evidence supporting each assessment. This transparency is designed to augment clinical decision-making, allowing medical professionals to verify AI-generated insights alongside their own evaluations. In benchmark evaluations across datasets for attention deficit hyperactivity disorder, Alzheimer’s disease, and autism spectrum disorder, 4DfCF consistently outperformed existing comparative models. On the Alzheimer’s Disease Neuroimaging Initiative dataset, the system achieved an F1 score of 96.28 percent. The research also demonstrated strong transfer learning capabilities, indicating that models pretrained on one neuroimaging cohort could rapidly adapt to new datasets with improved accuracy. This scalability points toward a future of reusable, specialized AI models for diverse neurological and neurodevelopmental conditions. Computational efficiency was a primary design objective. The main 4DfCF architecture contains approximately 10.34 million parameters, with a lightweight variant reduced to 4.18 million. These lean structures demand significantly fewer computational resources than larger contemporaries, making deployment on hospital servers and regional research infrastructure more viable. The model’s balanced architecture aligns with emerging standards for medical AI, prioritizing performance, interpretability, scalability, and operational efficiency. While current validation relies on benchmark datasets rather than live clinical environments, the findings establish a robust foundation for trustworthy computer-aided diagnostic systems. By combining high-fidelity pattern recognition with transparent reasoning and manageable resource requirements, 4DfCF advances the trajectory toward AI tools that assist, rather than supplant, medical experts in neurological screening and long-term patient monitoring.