Duplicate brain scans hidden in widely used medical datasets have been inflating the accuracy of AI tumor detection models, making them appear far more capable than they actually are in real clinical settings. A team of researchers at Northeastern University's Khoury College discovered this critical flaw and developed both a data-cleaning solution and an improved AI architecture to address it, achieving 98.5% accuracy while outperforming 13 baseline models.

What Is Data Leakage and Why Does It Matter in Medical AI?

The problem, known as data leakage, occurs when information from a model's training data accidentally appears in its testing data. In medical AI systems, this creates dangerously misleading results. A model may seem capable of diagnosing tumors with near-perfect accuracy on paper, but its performance may not hold up when presented with new patients in a real clinical setting.

"We found that the digital image libraries used worldwide to train brain tumor AI were full of duplicate images. AI models looked incredibly smart on paper, but they were actually memorizing leaked images," said Divya Chaudhary, associate teaching professor at Khoury College in Seattle and director of the Insight X Research Group.

Divya Chaudhary, Associate Teaching Professor at Khoury College of Computer Sciences

Chaudhary compared the problem to a student taking an exam after accidentally receiving half of the test questions in advance. The issue proved especially significant because medical imaging datasets are notoriously difficult to obtain due to privacy regulations, which may explain why researchers overlooked this problem for so long.

How Did Researchers Identify and Fix the Duplicate Data Problem?

Graduate students Dong Lu and Yu Zhang built an automated data-cleaning pipeline that identified and removed duplicate and near-duplicate MRI scans using file fingerprinting and image similarity analysis techniques. The team created a leakage-free benchmark dataset containing 3,522 unique MRI images, establishing a cleaner foundation for training trustworthy AI models.

After establishing the cleaner dataset, the researchers developed a new AI model called MTA-Swin, which builds on the existing Swin Transformer architecture commonly used for medical image analysis. The team modified the architecture by adding a multi-token attention mechanism that allows different parts of the model to exchange information more effectively, preserving fine-grained details that traditional approaches might miss.

Steps to Improve Medical AI Model Reliability

• Data Validation: Implement automated pipelines to detect and remove duplicate and near-duplicate images from training datasets before model development begins.
• Architecture Enhancement: Use multi-token attention mechanisms in vision transformers to enable different model components to communicate more effectively and preserve subtle diagnostic details.
• Explainability Integration: Apply visualization techniques like Grad-CAM to highlight which regions of medical images influenced the AI's decision, allowing physicians to verify the model is focusing on relevant features.

The researchers also focused on making the model more understandable for medical professionals. Using a visualization technique called Grad-CAM, the system can highlight the regions of an MRI scan that influenced its decision, allowing doctors to see whether the AI was focusing on the actual tumor or unrelated image features.

"Think of it like an expert radiologist. The model has multiple digital eyes communicating with each other, making sure fine-grained details are not missed," explained Chaudhary.

Divya Chaudhary, Associate Teaching Professor at Khoury College of Computer Sciences

The model demonstrated several promising capabilities in testing, perfectly identifying healthy brain scans while also showing high precision in detecting aggressive tumors. Importantly, the system was designed to operate without requiring expensive GPU (graphics processing unit) infrastructure, potentially making it more accessible for smaller clinics or other resource-limited providers.

Why Explainability Matters for Clinical Adoption

The researchers emphasize that AI is not meant to replace physicians, but rather to serve as a diagnostic aid. Explainability is critical for healthcare adoption because doctors will never trust an AI system if they cannot understand how it reaches its conclusions.

"AI should not and could not replace doctors. It gives doctors a second option. It can help them quickly review MRI images and provide additional information," noted Yu Zhang, computer science master's student at Khoury College.

Yu Zhang, Computer Science Master's Student at Khoury College of Computer Sciences

For Zhang and Lu, the project represented a major milestone as their first published research paper. Both said the experience taught them how to conduct research from the ground up, from reading scientific literature to designing experiments and responding to reviewer feedback. After developing the initial version in approximately three months, the team spent several additional months revising experiments and strengthening the paper before acceptance.

Chaudhary credits her students for leading much of the technical work and sees the project as an example of how careful engineering and responsible AI development can improve healthcare technology. The goal remains a more reliable, safer clinical diagnosis system that gives doctors better tools to make informed decisions.