Research Preview v2.1.0 --

Tri-Netra

Brain MRI tumor analysis with a cascade segmentation network, three independently-architected classifiers, and a evidence-grounded layered LLM radiology report.

Source on GitHub

97-99%

Classifier accuracy
CNN / Transfer / ViT on Kaggle test set

0.91

Segmentation micro-Dice
v3 UNet, BraTS 2020 test split

~30 ms

Inference latency
ONNX runtime, single 256x256 image

Zero

Validation
Evaluated on 246-sample OOD benchmark

Pipeline overview Click to expand the step-by-step methodology

1

Upload & normalize

2D MRI slice (PNG/JPG) is resized to 256x256, ImageNet-normalized for the segmentation backbone, and 224x224 for the classifiers.

->|

2

Cascade segmentation

Grayscale input is auto-detected and routed to the T1c specialist; multi-modal input goes to v3. 4-way TTA averaging + largest-component filter.

->|

3

3-classifier ensemble

CNN + ResNet50 transfer + ViT-hybrid run in parallel via ONNX. Epistemic = std across models, aleatoric = entropy of mean.

->|

4

Feature extraction

Deterministic radiology features: geometry, intensity, GLCM texture, morphology, mass effect, internal architecture, grade-evidence score.

->|

5

Layered LLM report

Pattern A polish, Pattern B citation-checked differential, Pattern C/D vision observer. Every LLM claim is validated against measured features.

BraTS 2020 (Bakas et al., 2018) LGG-MRI Segmentation (Buda et al., 2019) SMP UNet + ResNet34 (Iakubovskii, 2019) Grad-CAM (Selvaraju et al., 2017) Score-CAM / Occlusion (Wang et al., 2020) Llama 3.3 70B Instruct (Meta, 2024) Gemma 3 27B IT (Google DeepMind, 2025)

Run an analysis

Drop a single MRI here, or use Batch Upload above for many at once.

Drop MRI Scan Here

or click to browse files

Supports: PNG, JPG, JPEG (Max 50MB)

Analysis Model

Patient ID (Optional)

Scan Preview

MRI scan preview will appear here

File Name --

Dimensions --

File Size --

Analysis Results

Scan: SCAN-2024-0848 · Analyzed at 2:30 PM

Diagnosis Analyzing... AI model processing

Confidence --

Ensemble Rule -- Operating point: --

Estimated Volume -- cm³ Calculated from 2D area

Processing Time -- GPU Accelerated

AI Detector Summary

Decision rule: --
Mode: -- · --

Pattern Detector --

Compares regions to healthy brains

Score: -- · Alert level: --

Reconstruction Detector --

Tries to redraw the scan and flags struggles

Score: -- · Alert level: --

Tumor Outline Drawer --

A medical AI trained to outline tumors

Tumor area: -- pixels · Alert level: -- pixels

Asymmetry Detector --

Compares left vs right side of the brain

Score: -- · Alert level: --

Conformal Counterfactual Analysis

NOVEL

Per-voxel prediction sets with a provable (1 - α) coverage guarantee under causal interventions on the scanner protocol. Combines CONSeg-style conformal segmentation with CausalX-Net-style counterfactuals via weighted conformal prediction (Tibshirani et al. 2019). First unified framework for brain tumor MRI.

weighted-conformal-CF

Coverage target

--

Interventions

--

Most-disagreeing

--

Most-robust

--

Per-intervention table (q, abstain %, certified disagree %)

Intervention	q	Abstain %	Cert. disagree %	CF area px

Conformal Counterfactual Analysis NOVEL · RESEARCH HEAD

Novel research module (Gap B): voxelwise prediction sets with provable (1 - α) coverage guarantee under causal interventions on the scanner protocol. Combines CONSeg + CausalX-Net via weighted conformal prediction.
Awaiting first analysis with calibration artifacts. If you see this after running an analysis, the artifacts are still downloading from the model repo (cold-start ~60s); refresh and re-analyse to pull the populated panel.

Visualization

AI Insight Maps

Layperson-friendly · "where is the AI looking?"

Looks normal

Highly unusual

Interactive XAI Blend RED = MULTIPLE DETECTORS AGREE

Use the slider to blend the AI heatmap directly over the original MRI scan.

Heatmap Opacity

Pattern Detector

Checks each region against what healthy brains look like in a large self-supervised vision model.

Reconstruction Detector

A generative model tries to redraw each pixel from noise; this shows where it struggles most — usually unusual structure.

Asymmetry Detector

Compares the left and right sides of the brain. Highlights where one side differs from its mirror — common with unilateral lesions.

Inference Telemetry

Runtime: --
Total elapsed: --

Cascade Decision

Model used: --
Reason: --
Tumor area: --
Mean prob in mask: --

AI Radiology Report

Structured radiology-style report grounded in measured features, with a layered LLM enrichment pass (polish, differential expansion, visual co-observer). Every LLM claim is citation-checked or conflict-flagged against measurements.

Click "Generate Report" to run the layered LLM pipeline on this scan. The deterministic radiology report ships every time; the LLM pass adds polished prose, citation-checked differential diagnoses, and visual co-observations.

Tumor Segmentation

Pixel-level tumor boundary detection using Attention U-Net

Segmentation Model

Threshold 0.50

LLM Backend

Original MRI

Segmentation Mask

Overlay

Dice Score --

IoU --

Tumor Area --

Model Used -- --

AI Explanation

-- --

Impression

--

Overall Confidence

--

Why the Classifiers Ruled This Out

Vision LLM Reasoning (Pattern D - negative-case)

The vision model was shown the original MRI and asked to describe what visible features support the no-tumor verdict. Output is validated to ensure it does not contradict the verdict.

Structured Findings

Grade-Evidence Score

--

False-positive region analysis (debug)

Differential Diagnosis (citation-checked)

Visual Observations (LLM-co-observer)

Recommendation

--

Classifier Agreement

--

LLM Pass Status

Disclaimer

Not a medical diagnosis. Research / educational only.

Polished summary (verified prose, may equal Impression if LLM polish rejected)

Raw deterministic features (JSON)

Model Performance Comparison

Compare accuracy, speed, and performance metrics across all AI models

CNN (Fast)

Fast

Accuracy 85-90%

Inference Time ~0.5s

Parameters 2.3M

ROC AUC 0.87-0.92

Lightweight convolutional neural network optimized for speed. Best for real-time screening applications.

Transfer Learning

Balanced

Accuracy 89-93%

Inference Time ~1.2s

Parameters 25M

ROC AUC 0.91-0.95

Pre-trained ResNet architecture fine-tuned on medical imaging data. Optimal balance of speed and accuracy.

Vision Transformer

Accurate

Accuracy 95-98%

Inference Time ~2.5s

Parameters 86M

ROC AUC 0.96-0.99

State-of-the-art transformer architecture with attention mechanisms. Highest accuracy for critical diagnoses.

Performance Metrics Overview

CNN 85%

Transfer 91%

ViT 97%

Accuracy Comparison

Analysis Reports

View and manage historical scan reports and patient records

Scan ID	Patient ID	Date	Diagnosis	Confidence	Model
SCAN-2024-0847	PT-2024-0156	May 25, 2026	Tumor	96.2%	ViT
SCAN-2024-0846	PT-2024-0155	May 25, 2026	Normal	98.1%	Transfer
SCAN-2024-0845	PT-2024-0154	May 24, 2026	Tumor	94.7%	ViT
SCAN-2024-0844	PT-2024-0153	May 24, 2026	Normal	97.3%	CNN
SCAN-2024-0843	PT-2024-0152	May 23, 2026	Tumor	92.8%	Transfer

Total Scans

847

Tumor Positive

312

Normal

535

Avg. Confidence

95.4%

Tri-Netra AI

Tri-Netra AI

I am a Patient

I am a Doctor

97-99%

30ms

0.91

2.5k+

Welcome. Let's look at your scan.

Result

Risk Level

Recommended Next Steps

Have Questions?

Tri-Netra

Upload & normalize

Cascade segmentation

3-classifier ensemble

Feature extraction

Layered LLM report

Run an analysis

Batch Analysis

Drop MRI Scan Here

Scan Preview

Analysis Results

AI Detector Summary

Conformal Counterfactual Analysis

Tumor Boundary Refiner

Visualization

AI Insight Maps

Uncertainty Analysis

Robustness Score

Inference Telemetry

Cascade Decision

AI Radiology Report

Tumor Segmentation

Original MRI

Segmentation Mask

Overlay

Longitudinal Growth Timeline

AI Explanation

Impression

Overall Confidence

Why the Classifiers Ruled This Out

Vision LLM Reasoning (Pattern D - negative-case)

Structured Findings

Grade-Evidence Score

Differential Diagnosis (citation-checked)

Visual Observations (LLM-co-observer)

Model Disagreements (flagged conflicts)

Recommendation

Classifier Agreement

LLM Pass Status

Quality Warnings

Disclaimer

Model Performance Comparison

CNN (Fast)

Transfer Learning

Vision Transformer

Performance Metrics Overview

Analysis Reports

Total Scans

Tumor Positive

Normal

Avg. Confidence

Analyzing MRI Scan

Send Clinical Report