NEUROSCAN/4█
Multi-modal MRI brain tumour segmentation. Four-channel MRI ingestion (T1, T1ce, T2, FLAIR) with 31.5M-parameter Attention U-Net delineation of peritumoral edema, enhancing core, and necrotic tissue.
- Edema Halo (L2)
- WT 90.80% F1
- Enhancing Rim (L3)
- ET 82.76% F1
- Necrotic Core (L1)
- TC 85.23% F1
SCROLL TO ADVANCE GANTRY TO BRAIN ↓
MULTI-MODAL SLICE CONSOLE
3D CONTOUR STACK RECONSTRUCTION
Drag inside the CRT viewport to orbit the reconstructed brain mesh. Scrub STACK to watch 2D axial contours assemble into the 3D lesion.
SIGNAL PROCESSING & NEURAL ARCHITECTURE
31.5M ATTENTION U-NET
- [1]4 encoder-decoder levels with DoubleConv (Conv3×3 + BatchNorm + ReLU).
- [2]Additive Attention Gates filter skip connections to suppress healthy tissue noise.
- [3]Trained with Weighted Dice (1.2) + Focal Loss (1.5, γ=2.0) under FP16 AMP.
OSCILLOSCOPE CONVERGENCE TRACES
Trained 25 epochs on NVIDIA RTX Pro 6000 (55-case held-out validation split, TTA ×3 + calibrated sub-region thresholds).
| STAGE / REF | ARCHITECTURE & CONFIGURATION | WT F1 | TC F1 | ET F1 | MEAN F1 | BAL. ACC | VOX. ACC |
|---|
ETHICS, INTER-RATER LIMITS & CLINICAL SCOPE
Research prototype, not for clinical diagnosis. NEUROSCAN/4 is an academic and engineering demonstration trained on retrospective research scans (MICCAI BraTS 2020). It has not been cleared by the FDA, CE, or CDSCO and must never be used to guide patient care, biopsy targeting, or neurosurgical resection margins.
On glioblastoma MRI, board-certified neuroradiologists exhibit an inter-observer Dice agreement of ~82%–85% on the microscopic enhancing rim (ET). Evaluation numbers reported here reflect a 55-patient held-out split from the BraTS 2020 training archive.
While 2D axial slices with Test-Time Augmentation and 3D connected-component filtering achieve fast inference and high overlap, out-of-plane anisotropic spacing and scanner-specific coil artifacts require site-specific domain calibration before clinical translation.