Diffusion-enhanced AI framework segments and classifies brain tumors with 98.1% accuracy

Diagnosing brain tumors from magnetic resonance imaging (MRI) remains a difficult task, complicated by tumor heterogeneity, low contrast between tumor and healthy tissue, imaging artifacts and differences between tumor types. A new study published in Brain Informatics introduces a Promptable Diffusion-Enhanced Foundation Framework designed to perform both tumor segmentation and classification in a single, explainable pipeline, tested on the recent BRISC 2025 dataset.

The framework addresses a known weakness of today’s deep learning methods, which the authors note often lack explainability, tolerance to corrupted imaging conditions and the ability to interpret complex tumor structures. The system works in stages. First, a diffusion-based enhancement module improves the visibility of tumor boundaries and reduces imaging artifacts before any features are extracted. A pre-trained foundation encoder then extracts strong semantic representations from the enhanced images, and prompt-guided attention mechanisms adaptively localize tumor areas within the scans.

Unlike conventional pipelines that handle segmentation and classification separately, the framework performs both tasks jointly. Interpretation is supported by attention visualization and attention-based explanation maps, which are intended to make the model’s decisions more transparent to the user. According to the reported experimental results, the approach achieved a classification accuracy of 98.1%, a Dice score of 93.2% for segmentation, and an area under the curve (AUC) of 99.0%, which the authors describe as superior to conventional CNN and transformer-based comparison methods.

The study also reports that the framework demonstrated high robustness in the presence of imaging noise and low MRI contrast conditions. The authors argue that this robustness, combined with the built-in explainability features, would be important for clinically reliable and interpretable analysis of brain tumors. The article is published open access, with the version shared early as a peer-reviewed, accepted manuscript that is citable and carries a permanent DOI, subject to further edits before the final Version of Record.

Subject of Research: A deep learning framework combining diffusion-based MRI enhancement, foundation model features and prompt-guided attention for explainable brain tumor segmentation and classification.

Article Title: Promptable diffusion-enhanced foundation framework for explainable brain tumor segmentation and classification

Article References: Joseph, J. (2026). Promptable diffusion-enhanced foundation framework for explainable brain tumor segmentation and classification. Brain Informatics. https://doi.org/10.1186/s40708-026-00341-x

Image Credits: AI Generated

DOI: 10.1186/s40708-026-00341-x

Keywords: brain tumor, MRI, segmentation, classification, diffusion model, foundation model, explainable AI, attention mechanism, BRISC 2025, deep learning, medical imaging, Promptable