Two-step AI system boosts brain tumor detection from MRI scans

Researchers from the University of Sharjah say a new two-stage AI system can detect brain tumors in MRI scans and classify them into three types with higher accuracy than conventional approaches. The study, published in Healthcare Analytics, points to faster, more reliable screening tools, but the model still needs broader clinical validation before routine use.

Why it matters: - Brain tumor detection is time-sensitive, and faster screening can improve treatment outcomes. - The researchers say the new system could help clinicians identify tumors earlier and with greater reliability. - Brain tumors remain a major global health burden, causing an estimated quarter of a million deaths each year, according to the World Health Organization.

What happened: - Researchers from the University of Sharjah developed a two-step AI method for analyzing brain MRI scans. - The study was published in Healthcare Analytics. - The first step determines whether an MRI scan contains a tumor. - The second step classifies detected tumors as glioma, meningioma, or pituitary tumor. - The authors analyzed 7,023 MRI images covering three tumor categories and healthy cases. - The team used a unified preprocessing pipeline for all images. - The pipeline included resizing, intensity normalization and contrast equalization.

The details: - The study compared several deep learning approaches under the same clinical workflow and preprocessing rules. - Traditional convolutional neural network models performed well, but CNN-LSTM combinations did better. - Attention-enhanced architectures improved feature representation and classification accuracy by capturing more global patterns in MRI images. - The CNN-LSTM model with attention performed especially well in multi-class tumor classification. - The Vision Transformer model delivered the strongest results for binary detection of tumor versus non-tumor scans. - The system processed scans quickly, averaging 21.6 milliseconds for binary tasks and 17.6 milliseconds for multi-class tasks. - The authors say that speed supports potential near real-time clinical use. - The evaluation followed an image-level protocol designed for transparency and reproducibility. - The researchers note that full patient identifiers were not consistently available across source datasets, so strict patient-level separation could not be guaranteed.

Between the lines: - The study is not a new diagnostic device. It is a structured comparison of existing AI architectures for a clinically relevant task. - The results suggest different model families may be useful for different parts of the workflow, with Vision Transformers standing out for detection and CNN-LSTM models excelling at classification. - The strongest findings come from controlled dataset testing, not from prospective hospital deployment. - The authors acknowledge that publicly available MRI datasets may not fully reflect real patients, hospitals and imaging equipment.

What's next: - The researchers say more testing on larger and more diverse clinical datasets is needed before routine medical use. - Future work may add interpretability tools such as feature attribution and lesion-level visualization. - Those tools could help connect AI predictions with tumor pathology and increase clinician trust. - The authors also want to test robustness under low-resolution or highly variable MRI conditions. - Clinical validation will be the key step before the system can move from promising research to deployment.