🧠 Brats-3D — Brain Tumor MRI Analysis Platform

---

link : https://brats-3d.web.app/

An end-to-end AI-powered clinical platform for automated brain tumor segmentation using multi-modal MRI scans, built on the BraTS2020 benchmark dataset. The system accepts four MRI modalities as input and returns a detailed segmentation report identifying tumor sub-regions with clinical significance.

---

Dataset — BraTS2020

The Brain Tumor Segmentation (BraTS) dataset consists of multi-institutional, pre-operative MRI scans from 19 institutions. All scans are co-registered to a standard anatomical template, resampled to 1mm³ isotropic resolution, and skull-stripped.

Input Modalities

Modality	Full Name	Clinical Role
T1	T1-weighted (native)	Baseline brain anatomy, identifies normal tissue structure
T1CE (T1Gd)	T1-weighted post-contrast	Highlights active tumor using gadolinium contrast agent
T2	T2-weighted	Shows edema and infiltration, high signal in fluid regions
FLAIR	T2 Fluid Attenuated Inversion Recovery	Suppresses CSF signal, best for peritumoral edema detection

Segmentation Classes

Label	Class	Name	Description
1	NCR/NET	Necrotic and Non-Enhancing Tumor Core	Dead tissue at the center of the tumor. Appears dark on T1CE. Represents the oldest, most necrotic part of the tumor mass.
2	ED	Peritumoral Edema	Swelling around the tumor caused by fluid leakage. Best visible on T2/FLAIR. Does not contain active tumor cells but causes significant intracranial pressure.
4	ET	GD-Enhancing Tumor	Actively growing, invasive tumor region. Lights up on T1CE due to blood-brain barrier breakdown. Most clinically significant region for treatment planning and surgery.

Composite Tumor Regions (used in evaluation)

Region	Composed of	Clinical Meaning
Whole Tumor (WT)	ET + ED + NCR/NET	Total tumor extent including all affected tissue
Tumor Core (TC)	ET + NCR/NET	The core mass excluding surrounding edema
Enhancing Tumor (ET)	ET only	Active tumor, used for surgical and radiation targeting

---

Model Performance

Metric	Value	Description
Train Loss	0.3547	DiceCE combined loss on training set
Val Mean Dice	0.7155	Average Dice score across all three tumor classes
Val Dice — Class 1 (NCR/NET)	0.8073	Segmentation accuracy for necrotic core
Val Dice — Class 2 (ED)	0.8434	Segmentation accuracy for peritumoral edema
Val Dice — Class 3 (ET)	0.8261	Segmentation accuracy for enhancing tumor

Dice Score measures overlap between predicted and ground truth segmentation. A score of 1.0 means perfect overlap.

Training Configuration

Parameter	Value
Architecture	3D UNet
Dataset	BraTS2020
In Channels	4 (T1, T1CE, T2, FLAIR)
Out Channels	4 (background + 3 classes)
Optimizer	AdamW
Loss Function	DiceCELoss
Learning Rate	0.0001
Batch Size	2
Max Epochs	100
Early Stopping	Patience 15
Device	CUDA

---

System Architecture

Frontend (React + Vite)
        ↓ Auth0 JWT
FastAPI Backend  ──────────────────────→  Cloud Run (us-central1)
        ↓                                          ↓
GCS Signed URL                          Neon PostgreSQL
        ↓
Google Cloud Storage (brats-uploads)
        ↓ GCS paths
BentoML Inference Service  ────────────→  Cloud Run (us-central1)
        ↓
   3D UNet Model
        ↓
Segmentation Report + NIfTI Output

---

Tech Stack

AI & ML

• 3D UNet trained on BraTS2020
• BentoML for model serving and API
• MLflow for experiment tracking and model registry
• PyTorch with CUDA support

Backend

• FastAPI — async REST API
• Auth0 — JWT-based authentication
• SQLAlchemy + Alembic — ORM and database migrations
• Prometheus + Grafana — metrics and monitoring
• Redis — prediction result caching

Infrastructure

• Google Cloud Run — serverless container deployment
• Google Cloud Storage — large MRI file storage (bypasses 32MB Cloud Run limit)
• Neon PostgreSQL — serverless production database
• Firebase Hosting — frontend CDN deployment
• GitHub Actions — full CI/CD pipeline

---

CI/CD Pipeline

Every push to main automatically triggers:

1. Spin up isolated PostgreSQL test container
2. Run full test suite with pytest
3. Run Alembic migrations against production Neon database
4. Build and push FastAPI Docker image to GCR
5. Build and push BentoML Docker image to GCR
6. Deploy FastAPI to Cloud Run
7. Deploy BentoML to Cloud Run
8. Build React frontend with production environment variables
9. Deploy frontend to Firebase Hosting

---

File Upload Flow

Standard Cloud Run requests are limited to 32MB, which MRI NIfTI files easily exceed. The platform handles this with GCS signed URLs:

1. Frontend requests a signed upload URL from FastAPI
2. FastAPI generates a 15-minute PUT signed URL via GCS
3. Frontend uploads the file directly to GCS — bypassing Cloud Run entirely
4. Frontend sends only the GCS file paths to FastAPI for inference
5. BentoML downloads files from GCS, runs the 3D UNet, and returns results

---

---

API Endpoints

Method	Endpoint	Auth	Description
GET	/health	No	Health check
GET	/upload/signed-url	Yes	Get GCS signed URL for direct MRI file upload
POST	/predict/	Yes	Submit GCS file paths and run tumor segmentation
GET	/users/me	Yes	Get authenticated user profile
POST	/login	No	Auth0 token exchange
GET	/metrics	No	Prometheus metrics scrape endpoint

---

Author

Waseem Almazrua — LinkedIn