About — BraTS Viewer
Interactive 3D visualization of glioblastoma segmentation on the BraTS 2023 GLI challenge.Independent research
Built as a companion to my **independent research report** (no academic supervision, no commission) — the viewer shows not just a model output but the full therapeutic chain (diagnosis → surgery → radiotherapy) that a clinical AI has to plug into.
1. Clinical context
Glioblastoma (GBM) is the most aggressive primary brain tumor in adults, with a median survival of 12–15 months under the current standard-of-care. Accurate segmentation of the tumor subregions (enhancing tumor, necrotic core, peritumoral edema) drives three downstream decisions : surgical planning (maximal safe resection), radiotherapy targeting (GTV → CTV → PTV expansion), and longitudinal response assessment (RANO criteria).
2. Dataset — BraTS 2023 GLI
The BraTS 2023 Adult Glioma (GLI) subchallenge provides 1,251 patients with four co-registered MRI sequences (T1, T1ce, T2, FLAIR) and expert-annotated segmentation masks for necrotic core, enhancing tumor and peritumoral edema. All volumes are skull-stripped and resampled to 1 mm³ isotropic in SRI24 atlas space.
3. Model — nnU-Net v2 + MedNeXt-B + Distance Map
The segmentation used here is produced by MedNeXt-B (Roy et al., MICCAI 2023) trained through nnU-Net v2 with an auxiliary distance-map regression head. The distance map branch predicts the signed distance to the tumor surface (Ma et al., 2020), which acts as a self-decaying regularizer nudging the main head toward boundary-coherent outputs.
Ablation of the loss coefficient λ_dist on fold 0 (300 epochs, seed=42) shows a flat-ish plateau around λ = 0.1 (val Dice 0.9077), matching the theoretical gradient-equilibrium estimate of 2 % of the dominant Dice+CE loss magnitude. In practice the auxiliary MSE decays naturally from 3.8 % of the segmentation gradient at epoch 0 to 0.2 % by epoch 150+, acting as an early regularizer that fades out as the model converges.
4. Downstream — surgery & radiotherapy simulation
The viewer embeds two educational simulations that show how a segmentation plugs into the therapeutic pipeline :
Surgery mode
8-phase animation : plan entry point (burr hole 14 mm Ø, biased toward convex superior cortex), drill craniotomy, lift bone flap, open dura, resect tumor, close dura with artificial substitute, replace bone flap with titanium plates + 3 screws. Parameters are realistic (pneumatic drill, 60k rpm) but the viewer is educational — not for clinical use.
Radiotherapy mode — Stupp protocol
Standard-of-care fractionation : 60 Gy total / 30 fractions of 2 Gy, delivered 5 days per week over 6 weeks, with concomitant temozolomide. The viewer renders 5 IMRT beams (6 MV photons, λ ≈ 2.07 × 10⁻¹³ m, ν ≈ 1.45 × 10²¹ Hz) converging on the tumor, with the GTV → CTV (+18 mm) → PTV (+5 mm setup margin) dose shells shown as concentric surfaces.
5. Architecture of this viewer
The front-end is Next.js 15 + React + Three.js (via @react-three/fiber) compiled as a 100 % static export. Everything the viewer needs — meshes, 2D slices, defect counts, dose shells, surgical plans — is pre-computed once on the tower (i7 E-cores + RTX PRO 6000) and served as flat files from S3 + CloudFront. Zero backend runs in production, so there is no cold-start latency when a reader opens the page.
- Brain mesh decimated to ~20 % triangles (quadric-edge simplification) for fast first paint.
- 2D slice contours carry both the vector loops and a base64-PNG mask (~11× smaller than nested JSON arrays).
- Every JSON asset is pre-compressed at build time as
.br(brotli q=11) and.gz(q=9), served via CloudFront content-negotiation. - Mesh and slice assets cached as
immutable, max-age=1y; HTML revalidated every 60 s so content updates propagate quickly.
6. References
- Baid et al. The RSNA-ASNR-MICCAI BraTS 2021 benchmark. arXiv:2107.02314.
- Roy et al. MedNeXt: Transformer-driven scaling of ConvNets for medical image segmentation. MICCAI 2023.
- Isensee et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods 2021.
- Ma et al. Distance map loss penalty for boundary-aware segmentation. MIDL 2020.
- Stupp et al. Radiotherapy plus concomitant and adjuvant temozolomide for glioblastoma. NEJM 352, 987–996 (2005).
Educational / research tool — not a medical device. All segmentations shown are from the public BraTS 2023 GLI training set and have been re-predicted by an unregulated research model.