Rapid AI System Detects and Classifies Brain Tumor Subtypes on MRI

Early detection and accurate subtype classification on magnetic resonance imaging are central to treatment planning, yet performance can vary across tools and sites. Faster analysis is also needed to support time-sensitive decisions in neuro-oncology. To help address this challenge, researchers have developed an artificial intelligence system that detects tumors on brain MRI and then classifies them by type.

Developed by scientists at the University of Sharjah, the approach uses a two-step workflow that mirrors clinical decision-making. The first step screens each scan for the presence of a lesion. If a lesion is detected, the second step assigns one of three categories: glioma, meningioma, or pituitary tumor. The study reports higher accuracy than conventional systems when advanced deep learning architectures are applied within this framework.

The method standardizes image handling before model inference. A unified preprocessing pipeline resizes images to a fixed resolution, normalizes intensity, and equalizes contrast to reduce site-specific variation. Feature extraction is performed with a pre-trained convolutional neural network (CNN). These features are modeled with long short-term memory (LSTM) layers, and an attention module refines global context for the final decision.

Model comparisons showed complementary strengths. An attention-enhanced CNN–LSTM achieved particularly strong performance for multiclass subtype classification. A Vision Transformer model delivered the highest performance on the binary tumor-versus-nontumor task. The evaluation followed a clearly defined image-level protocol applied to 7,023 MRI images spanning three tumor types and nontumor scans. Because some source datasets lacked patient identifiers, strict patient-level separation could not be guaranteed.

The system was computationally efficient, averaging 21.6 milliseconds per scan for tumor detection and 17.6 milliseconds for subtype classification, supporting near–real-time workflows. The authors note that reliance on publicly available datasets may limit generalizability, and they call for testing on larger, more diverse clinical cohorts. They also highlight the value of interpretability methods to align model outputs with tumor pathology. Findings were published in Healthcare Analytics.

“The experiments show that model performance improves when progressing from standard CNNs to CNN with LSTM and then to attention-enhanced architectures. Across both stages, attention mechanisms improved feature representation and classification accuracy by capturing more informative global patterns in MRI images,” said the researchers.

“The system showed efficient inference times, averaging 21.6 ms for binary and 17.6 ms for multiclass tasks, supporting its suitability for near-real-time clinical applications. Nonetheless, evaluating the model's robustness under low-resolution or highly variable MRI conditions remains an important avenue for future work,” stated the authors.

Related Links
University of Sharjah