Brain MRI Decision Support Project
Transparent intelligence for brain MRI review.
NeuroSight AI combines tumor segmentation, multi-class diagnosis, explainability overlays, and confidence-aware safeguards in one calm, research-focused workflow.
I. Overview
Why NeuroSight was built
Clinical Objective
NeuroSight AI is a brain MRI analysis assistant designed for clinician-in-the-loop research workflows. It reduces review friction by unifying segmentation, classification, and visual evidence in a single decision surface.
MVP Capabilities
- 2D slice-level segmentation with UNet++
- 4-class diagnosis with EfficientNetV2-S
- Grad-CAM and Integrated Gradients attribution maps
- Entropy-based uncertainty and rejection safeguards
II. Architecture
Pipeline from scan to case verdict
- 01 Batch MRI slices enter through the React review console
- 02 FastAPI coordinates preprocessing and model execution
- 03 UNet++ extracts lesion masks per image slice
- 04 EfficientNetV2-S outputs calibrated class probabilities
- 05 Aggregation + XAI generate a case-level interpretation
Backend
FastAPI, SQLAlchemy, SQLite, PyTorch model serving
Frontend
React, Vite, Tailwind UI with clinical viewer controls
Safety logic
Temperature scaling, uncertainty entropy, rejection thresholds
III. Live Result
Real NeuroSight output sample
What this demonstrates
This image captures the deployed inference workflow in action, including processed MRI context and generated visual outputs used for interpretation support.
It represents the production-style artifact format returned by the backend pipeline for downstream clinical review and auditing.
IV. Evidence
Model training and validation evidence
V. Roadmap
Next-stage evolution
Volumetric data
Move from 2D image slices to NIfTI 3D volume ingestion and preprocessing.
3D model stack
Evaluate 3D U-Net or V-Net segmentation with 3D classifiers.
Clinical viewer expansion
Introduce multi-planar rendering and richer uncertainty diagnostics.