-
-
Gradio powered ML Demo
-
Integrated Diagnosis (MRI + clinical assessment)
-
About~
-
Model card
MirAI (未来): Protecting the Future of Brain Health
🌟 Inspiration
Alzheimer’s Disease is a progressive journey that often remains hidden until significant damage is done. The inspiration for MirAI (Japanese for Future) came from a critical gap in current diagnostic workflows: the disconnect between structural neuroimaging and cognitive behavioral testing. We wanted to build a system that doesn't just provide a diagnosis, but offers a consensus—combining what the brain looks like (MRI) with how the brain performs (Clinical scores). Our mission is to "Protect the Future" of patients by enabling earlier, more transparent AI-assisted detection.
🧠 How We Built MirAI
We engineered a Parallel Dual-Expert System using the extensive OASIS-3 longitudinal dataset.
The Structural Expert (3D MRI): We implemented a 3D DenseNet-121 architecture to process volumetric T1-weighted MRI scans (128*128*128). Unlike 2D models, this captures the spatial relationships of brain atrophy across the hippocampus and ventricles.
The Functional Expert (Clinical MLP): A Multi-Layer Perceptron (MLP) was trained on UDS (Uniform Data Set) clinical records. It analyzes high-impact cognitive biomarkers including MMSE (Mini-Mental State Exam) scores and Clinical Dementia Rating (CDR) values.
Data Alignment Pipeline: We developed a custom time-alignment algorithm to synchronize MRI scans with the closest clinical visit data, ensuring that the "structural" and "functional" snapshots of the patient matched in time.
Clinician Dashboard: The interface was built using Gradio with a premium, scrollable UI/UX, designed to provide physicians with a side-by-side comparison of structural and cognitive findings.
🚀 Challenges We Faced
- Computational Complexity: Processing 3D volumes is resource-intensive. We had to optimize our pipeline for an Nvidia DGX-1, utilizing mixed-precision training and efficient data loading to handle 128^3 voxels without memory overflows.
- Data Disparity: With over 8,000 clinical records but only 4,000 MRI scans, aligning the data without introducing "leakage" or bias required rigorous subject-level splitting.
- Architecture Pivot: Initially, we explored feature-level fusion, but since the access the GPU server was limited , a Parallel Architecture was superior. It allows doctors to see where the discrepancy lies if the models disagree, rather than being forced to trust a single "black box" score.
📚 What We Learned
- Medical Context Matters: We learned that high accuracy isn't enough; medical AI must be interpretable. By presenting two separate "expert reports," we provide a more robust decision-support tool.
- Data Engineering is 80% of the Work: Building the time-alignment scripts was just as important as the neural network architecture itself.
- 3D Volumetric Insight: We discovered how 3D convolutions can identify subtle atrophy patterns that are often missed in traditional 2D slice analysis.
📊 Performance Summary
We achieved a balanced performance that leverages the strengths of both modalities:
- Clinical Accuracy: 87%
- MRI Structural Accuracy: 84%
- System Consensus Accuracy: ~90%
🔮 Future Enhancements: The MirAI Fusion Roadmap
Our current parallel architecture is only the beginning. We are moving toward a Unified Multi-Modal Fusion Engine that will integrate:
- Metabolic Imaging (PET Scans): Incorporating Amyloid/FDG-PET data to detect metabolic changes before structural atrophy occurs.
- Fluid Biomarkers: Integrating CSF (Cerebrospinal Fluid) and blood-based protein markers (Tau/Amyloid-beta) for biochemical validation.
- Triple-Modality Fusion: Developing a late-fusion transformer architecture where MRI + PET + Biomarkers + Clinical Scores are weighted mathematically to provide a singular, high-confidence "Neuro-Health Index."
MirAI (未来) — Utilizing AI to preserve the memories of tomorrow.
Log in or sign up for Devpost to join the conversation.