Web Preview vs. Full Application Architecture
This deployment is an optimized lightweight web preview of the complete NeuroFramework ecosystem:
Web Preview (This Site)
- Hosted on Vercel Edge CDN for instant global access.
- Zero server-side neural processing overhead on Vercel.
- Pre-computed 50-slice multi-planar tensors for 9 ABIDE-I cohort subjects.
- Interactive tri-planar 60fps scrubbing & CBAM attention mapping.
- Single-page medical report generation.
Full Application (NeuroFramework-ai)
- Complete autonomous PyTorch 3D deep learning pipeline.
- Live raw NIfTI (
.nii, .nii.gz) file ingestion & parsing.
- Adaptive Otsu brain masking & N4 bias field correction.
- Live 3-Stream Conv3D-CBAM GPU neural inference.
- Full model training scripts, 5-fold CV & focal loss.
Access NeuroFramework-ai on GitHub ↗
Peer-Reviewed Reference
Hammash, N. M., & Younis, M. C. (2026). A Hierarchical Multi-View Deep Learning Framework for Autism Classification Using Structural and Functional MRI. MDPI Journal of Imaging, 12(3), 109.
Multi-Site Validation Benchmarks (ABIDE-I, N=395)
| Cohort / Protocol |
Resolution |
Peak Fold |
5-Fold Average |
| Multi-Site (NYU + UM_1 + USM) |
50 Slices @ 224x224 HD |
75.95% (Fold 1) |
66.84% |
| Single-Site (NYU Alone, N=184) |
50 Slices @ 128x128 |
75.00% |
69.59% |
Architectural Mechanics
- 3-Stream Conv3D: Alternating 3x3x3 and 5x5x5 kernels with residual skip connections.
- 3D CBAM: Dynamic channel attention and 3D spatial attention map extraction.
- Adaptive Focal Loss: Binary focal loss (γ = 2.0, α = 0.25) preventing class collapse.