Student Research Project

This project was built by a high school student to learn how deep learning models for brain MRI work and where they fall short. You can upload an image to explore regions a model flags and, for classification models, the tumor type it predicts. Reviewing real uploads shows the models frequently fail; this is not a clinical tool or a substitute for a radiologist reading the full study.

The Original Detection model moved to RF-DETR (Medium) in the August 20, 2025 v2.0.0 update. On January 27, 2026, the v3.0.0 model update added YOLO26 object detection and instance segmentation models that predict glioma, meningioma or pituitary tumor types. These experiments test different kinds of output: boxes, boxes with a predicted type, and pixel outlines with a predicted type.

Reviewing uploaded scans, especially ones flagged with "Report Incorrect", shows frequent real-world failures: a model may miss a tumor, flag normal tissue or predict the wrong type. Strong development metrics did not transfer to real uploads. Data diversity and evaluation design matter as much as architecture; there is no measured real-world accuracy yet.

Our planned goals include more varied multi-site and multi-sequence data, scanner and image-quality augmentation, consistent preprocessing, patient-level splits and external testing, confidence calibration and learning from reviewed reports. Today, only the "Report Incorrect" button and admin review of reports exist. Read what we've learned and why real-world results are lower.

This platform is intended strictly for research and educational use. It was developed to:

• Demonstrate the use of artificial intelligence in medical imaging analysis
• Encourage student and academic exploration into computer vision and health data
• Support non-clinical learning, experimentation, and innovation in the AI-healthcare space