🚧 Status: In Active Development (Not Production Ready Yet)
CerebraScan AI is an end-to-end multimodal MRI analysis system for brain tumor segmentation, volumetric analysis, radiogenomic inference, and automated clinical reporting, while integrating production-grade MLOps, data versioning, and deployment tooling.
It combines:
• Deep Learning (U-Net & nnU-Net)
• Radiogenomics
• 3D Visualization
• Longitudinal Tracking
• LLM-based Reporting
• DVC + S3 Data Management
• MLflow Experiment Tracking
• Dockerized Deployment
to support neuro-oncology research and clinical workflows.
This system targets real-world neuro-oncology use-cases through the following objectives:
- Volumetric 3D Tumor & Edema Analysis using normalized T1, T1CE, T2, and FLAIR modalities.
- Color-Coded Sub-Region Annotation for edema, NET, ET, and background.
- Intracranial Tumor Classification (Glioma, Meningioma, Pituitary, Schwannomas, Medulloblastoma, Ependymoma).
- 2D→3D Fly-Through Visualization with grayscale or RGB rendering.
- Longitudinal Delta Tracking for growth/shrinkage velocity and treatment response.
- Surgical No-Go Zone Identification to analyze anatomical adjacency and risk.
- Biopsy Site Recommendation via hotspot-based heatmaps for high-grade tissue sampling.
- Radiogenomic Virtual Biopsy predicting IDH mutation & MGMT methylation.
- Automated LLM-Based Reporting that consolidates all patient findings into structured radiology-style documentation.
Supported models:
• U-Net (TensorFlow/Keras) — baseline 2D/3D segmentation
• nnU-Net (PyTorch) — auto-configured SOTA clinical segmentation
nnU-Net is used as the default segmentation engine due to its automated preprocessing, architecture tuning, and training pipeline.
Post-segmentation, extracted radiomics + intensity features can be used for:
• Tumor type classification
• Low-grade vs high-grade prediction
Using TCGA-GBM, the system aims to non-invasively predict:
• IDH mutation status
• MGMT promoter methylation
3D reconstruction enables:
• ET / NET / Edema volume extraction
• Ratio & spread metrics
• Treatment response quantification
Supports:
• 2D slice overlays
• 3D surface rendering
• RGB segmentation overlays
• Fly-through mode
For multi-session scans of the same patient, the system computes:
• Tumor growth rate
• Edema evolution
• Volumetric velocity maps
• Clinical progression markers
• Multimodal MRI segmentation
• Clinical 3D volumetric mapping
• Sub-region color coding
• Tumor-type classification
• Radiogenomic virtual biopsy
• Biopsy hotspot recommendation
• Surgical risk/no-go analysis
• Longitudinal progression tracking
• DICOM/NIfTI pipeline
• MLOps-ready architecture
• Automated structured reporting
BraTS 2023
• Task: Tumor sub-region segmentation
• Modalities: T1, T1CE, T2, FLAIR
• Labels: Edema, NET, ET
TCGA-GBM
• Task: Radiogenomics + classification
• Labels: IDH, MGMT, survival metadata
• U-Net (Keras/TensorFlow)
CerebraScan AI includes a production-grade data & experiment workflow using:
• DVC for dataset versioning
• AWS S3 for remote storage
• MLflow for tracking experiments & metrics
• Docker for deployment
• Git branching for parallel experimentation
This transforms the project from a notebook-based ML experiment into a scalable MLOps pipeline for research labs & hospitals.
Admin (CPU)
│
│ dvc push (write)
▼
AWS S3 Bucket
▲
│ dvc pull (read)
│
GPU Machines (Training/Inference)
│
└── git push (experiment branches)
▼
Git Remote (GitHub/GitLab)
| Component | Tool |
|---|---|
| Data Versioning | DVC |
| Remote Storage | S3 |
| Experiment Tracking | MLflow |
| Access Control | IAM |
| Model Training | GPU Machines |
| Formats | NIfTI / DICOM |
| Deployment | Docker |
• S3 stores sensitive dataset artifacts
• DVC manages version references
• Git stores only metadata & code
• IAM restricts write-access to admin
• If used in a team, teammates have read-only S3 access
This project is maintained by Shardul Salodkar.