MRI Tumor Detection AI

Help Center

Learn how to explore this student research project and understand its limits

Quick Start Guide

1. Choose Model

Select an AI model based on your analysis needs

2. Upload

Select your MRI scan image (DICOM, PNG, or JPEG)

3. Review

View flagged regions, predicted types and model confidence

4. Assess

Use symptom checklist to evaluate consistency

Frequently Asked Questions

Choose a model based on the kind of output you want to explore. Each model does a different task:

Tumor boxes
Original Detection

RF-DETR (Medium), trained on DWI, draws boxes around regions it flags as possible tumor.

Output: Boxes with one "tumor" class; nothing flagged does not mean the scan is normal.

Boxes + tumor type
Tumor Classification (Object Detection)

YOLO26 object detection draws boxes and predicts glioma, meningioma or pituitary for flagged regions.

Output: Boxes with a predicted tumor type, not a diagnosis.

Outlines + tumor type
Tumor Classification (Segmentation)

YOLO26 instance segmentation draws pixel outlines and predicts glioma, meningioma or pituitary for flagged regions.

Output: Pixel outlines with a predicted tumor type, not a measured tumor boundary.

Supported File Types:

.dcm (DICOM)
.png
.jpg/.jpeg

Maximum file size: 5MB

1

Click "Choose File" or drag and drop

You can either click the upload button or drag your file directly onto the upload area.

2

Select your MRI scan file

Choose a brain MRI scan in one of the supported formats. The image should clearly show brain tissue.

3

Wait for upload confirmation

The system will validate your file and show a preview once successfully uploaded.

Important Reminder

This tool is for education and research only, not for diagnosis. A qualified radiologist must interpret a full study.

Model Output

  • Bounding boxes: Red rectangles around regions the model flags as possible tumor
  • Confidence scores: The model's own scores, not the chance that a result is correct
  • Label annotations: "Tumor" labels with model confidence percentages

Processing Time

Analysis typically takes 10-30 seconds depending on image size and complexity.

Download Options

You can download both the original image and the annotated version with detection boxes for comparison.

Educational Tool Only

The symptom checklist compares model output with selected symptoms for this research demo; it is not clinically validated or a diagnostic tool. Consult a qualified clinician about medical concerns.

What is the Symptom Consistency Assessment?

When using classification models, a symptom checklist appears based on the predicted tumor type. This feature calculates a "Combined Consistency Score" from model confidence and selected symptoms.

How the Score is Calculated

Formula: Score = (55 x AI Confidence) + (Symptom Points)

  • AI Confidence contributes up to 55 points (55% weight)
  • Symptom checkboxes contribute up to 45 points (45% weight)
  • Symptom weights are illustrative values chosen for this research demo; they are not clinically validated
  • Maximum possible score is 100

Tumor-Specific Symptoms

Glioma

Headaches, seizures, cognitive changes, motor weakness, speech difficulties

Meningioma

Headaches, vision changes, hearing loss, limb weakness, memory problems

Pituitary

Vision changes, hormonal symptoms, fatigue, unexplained weight changes

When Multiple Regions Are Flagged

If the model flags multiple regions, the symptom checklist uses the prediction with the largest model confidence score. An information banner indicates which predicted tumor type the checklist is based on.

High model confidence

Confidence Score:
85.7%

The model is very confident about this flagged region, but it can still be wrong.

Medium model confidence

Confidence Score:
58.3%

The model is uncertain about this flagged region. Its score cannot confirm what the image shows.

Understanding Confidence Scores

80-100%:Very confident - can still be wrong
60-79%:Fairly confident - can still be wrong
40-59%:Uncertain - needs review
Below 40%:Low confidence - often a false flag

Confidence is the model's own score, not the chance that the result is correct.

The published numbers (for example 97.9% mAP@50 for Original Detection) are development metrics. Each was measured on a held-out validation split of that model's own development dataset, which comes from the same curated collection as its training images. Real uploads are different: they come from other hospitals and scanners, use different MRI sequences (Original Detection was trained on DWI), are often screenshots, phone photos or crops, and include patients and conditions the models never saw. The models also look at a single 2D slice instead of a full 3D study, and their confidence scores are not calibrated probabilities. Because of this, real-world performance is substantially lower, and reviewing uploaded scans shows the models frequently fail. No real-world accuracy has been measured yet. The Model Transparency page explains this in detail.

AI Model Information

  • Uses RF-DETR (Medium) and YOLO26 computer vision models
  • Trained on medical imaging datasets
  • Explores detection, classification and segmentation of brain MRI images

System Limitations

  • Not certified for clinical diagnosis
  • Performance may vary with image quality
  • Original Detection was trained on DWI; other sequences can look different
  • Frequently misses tumors or flags normal tissue on real uploaded scans

Upload Issues

  • Ensure file is under 5MB
  • Check file format (DICOM, PNG, or JPEG only)
  • Try refreshing the page and uploading again

Processing Delays

  • Large files may take longer to process
  • High server load can cause delays
  • Wait at least 60 seconds before refreshing

No Detections Found

  • No regions flagged does not mean the scan is normal; models can miss tumors
  • Ensure the image clearly shows brain tissue
  • Try a different image or scan slice

Still Need Help?

Can't find what you're looking for? Email the project developer for help.

Email Support: mri.tumor.detection.ai@gmail.com

Available AI models and when to use them

Choose a model based on the kind of output you want to explore. Each model does a different task: Tumor boxes. Original Detection. RF-DETR (Medium), trained on DWI, draws boxes around regions it flags as possible tumor. Output: Boxes with one "tumor" class; nothing flagged does not mean the scan is normal. Boxes + tumor type. Tumor Classification (Object Detection). YOLO26 object detection draws boxes and predicts glioma, meningioma or pituitary for flagged regions. Output: Boxes with a predicted tumor type, not a diagnosis. Outlines + tumor type. Tumor Classification (Segmentation). YOLO26 instance segmentation draws pixel outlines and predicts glioma, meningioma or pituitary for flagged regions. Output: Pixel outlines with a predicted tumor type, not a measured tumor boundary.

How to upload MRI images

Supported File Types: .dcm (DICOM), .png, .jpg/.jpeg. Maximum file size: 5MB. Click "Choose File" or drag and drop. You can either click the upload button or drag your file directly onto the upload area. Select your MRI scan file. Choose a brain MRI scan in one of the supported formats. The image should clearly show brain tissue. Wait for upload confirmation. The system will validate your file and show a preview once successfully uploaded.

What to expect from the AI output

Important Reminder. This tool is for education and research only, not for diagnosis. A qualified radiologist must interpret a full study. Model Output. Bounding boxes: Red rectangles around regions the model flags as possible tumor. Confidence scores: The model's own scores, not the chance that a result is correct. Label annotations: "Tumor" labels with model confidence percentages. Processing Time. Analysis typically takes 10-30 seconds depending on image size and complexity. Download Options. You can download both the original image and the annotated version with detection boxes for comparison.

Understanding the Symptom Consistency Assessment

Educational Tool Only. The symptom checklist compares model output with selected symptoms for this research demo; it is not clinically validated or a diagnostic tool. Consult a qualified clinician about medical concerns. What is the Symptom Consistency Assessment? When using classification models, a symptom checklist appears based on the predicted tumor type. This feature calculates a "Combined Consistency Score" from model confidence and selected symptoms. How the Score is Calculated. Formula: Score = (55 x AI Confidence) + (Symptom Points). AI Confidence contributes up to 55 points (55% weight). Symptom checkboxes contribute up to 45 points (45% weight). Symptom weights are illustrative values chosen for this research demo; they are not clinically validated. Maximum possible score is 100. Tumor-Specific Symptoms. Glioma: Headaches, seizures, cognitive changes, motor weakness, speech difficulties. Meningioma: Headaches, vision changes, hearing loss, limb weakness, memory problems. Pituitary: Vision changes, hormonal symptoms, fatigue, unexplained weight changes. When Multiple Regions Are Flagged. If the model flags multiple regions, the symptom checklist uses the prediction with the largest model confidence score. An information banner indicates which predicted tumor type the checklist is based on.

Sample predictions and interpretation

High model confidence. Confidence Score: 85.7%. The model is very confident about this flagged region, but it can still be wrong. Medium model confidence. Confidence Score: 58.3%. The model is uncertain about this flagged region. Its score cannot confirm what the image shows. Understanding Confidence Scores. 80-100%: Very confident - can still be wrong. 60-79%: Fairly confident - can still be wrong. 40-59%: Uncertain - needs review. Below 40%: Low confidence - often a false flag. Confidence is the model's own score, not the chance that the result is correct.

Why are my results different from the published model metrics?

The published numbers (for example 97.9% mAP@50 for Original Detection) are development metrics. Each was measured on a held-out validation split of that model's own development dataset, which comes from the same curated collection as its training images. Real uploads are different: they come from other hospitals and scanners, use different MRI sequences (Original Detection was trained on DWI), are often screenshots, phone photos or crops, and include patients and conditions the models never saw. The models also look at a single 2D slice instead of a full 3D study, and their confidence scores are not calibrated probabilities. Because of this, real-world performance is substantially lower, and reviewing uploaded scans shows the models frequently fail. No real-world accuracy has been measured yet. The Model Transparency page explains this in detail.

Technical specifications and limitations

AI Model Information. Uses RF-DETR (Medium) and YOLO26 computer vision models. Trained on medical imaging datasets. Explores detection, classification and segmentation of brain MRI images. System Limitations. Not certified for clinical diagnosis. Performance may vary with image quality. Original Detection was trained on DWI; other sequences can look different. Frequently misses tumors or flags normal tissue on real uploaded scans.

Troubleshooting common issues

Upload Issues. Ensure file is under 5MB. Check file format (DICOM, PNG, or JPEG only). Try refreshing the page and uploading again. Processing Delays. Large files may take longer to process. High server load can cause delays. Wait at least 60 seconds before refreshing. No Detections Found. No regions flagged does not mean the scan is normal; models can miss tumors. Ensure the image clearly shows brain tissue. Try a different image or scan slice.