Release of Neuroflux v1.0: An Open-Source AI Model for Brain Tumor Segmentation

Demo Video of Neuroflux v1.0

We are thrilled to announce the official release of Neuroflux v1.0 on 24 February 2025. Our cutting-edge efficient AI-assisted brain tumor segmentation model is now available as an open-source project on GitHub. Neuroflux represents a significant step forward in neuro-oncology imaging, combining 3 critical aspects for clinical applicability: 

1. Enhanced accuracy and precision 

2. Unprecedented computational efficiency

3. Clear explainability of results

What is Neuroflux?

Neuroflux is a multimodal AI model built on a customised U-Net architecture specifically designed for glioblastoma multiforme (GBM) segmentation. With a low median overall survival rate of just 15 months, early intervention is critical to improving survival rates.

Designed to address a critical challenge in current medical practice and amongst existing AI models, Neuroflux bridges the gap between manual anatomical segmentation and an efficient, reliable AI model. Ultimately, Neuroflux also broadens the clinical applicability and accessibility in low-resource hospital environments.

The Neuroflux v1.0 model pipeline [Source: International Journal of High School Research, DOI: 10.36838/v7i6.64]

Beyond its architecture, Neuroflux incorporates a structured pipeline that enhances its overall clinical usability. Beginning with preprocessing, raw patient MRI data is cleaned, standardized, and balanced through augmentation to ensure robust learning across diverse tumor morphologies. The model then extracts critical features such as edge detection, textural patterns, and cellular morphology before passing through the deep learning backbone for tumor segmentation and risk assessment. The final layer of Grad-CAM–based explainability overlays generate heatmaps to highlight the specific regions influencing Neuroflux's predictions, while clinician validation and a feedback mechanism enable iterative refinement of the model. This end-to-end process ensures Neuroflux is not only accurate but also transparent, sustainable, and adaptable to real-world hospital workflows.

Release Highlights

Neuroflux introduces several groundbreaking features that distinguish it from existing segmentation models:

1. Lightweight Implementation

Optimised for computational efficiency without sacrificing accuracy

2. Multimodal MRI Analysis

Leverages both T1-contrast enhanced and FLAIR sequences for comprehensive GBM analysis

3. Grad-CAM Enhanced Visualisation

Provides real-time interpretability through clear tumour heatmaps

4. Custom Hybrid Loss Function

Combines Categorical Cross-Entropy with Dice Loss to improve segmentation accuracy

5. Adaptive Learning Rate & Dropout Optimisation

Ensures overall stability and prevents overfitting

Performance Metrics

Metrics	Scores
Pixelwise Accuracy	0.99
Mean Intersection over Union	0.7067
Precision	0.9922
Sensitivity	0.9873
Specificity	0.9973
Dice Coefficient (Necrotic)	0.445
Dice Coefficient (Edema)	0.6018
Dice Coefficient (Enhancing)	0.6564

These exceptional metrics were achieved while also significantly improving computational efficiency and clinical interpretability. Our high-precision and efficient modified architecture demonstrates the potential for lightweight and interpretable models to enhance diagnostic reliability and treatment planning. While our Dice coefficients across necrotic, edema, and enhancing subregions remained competitive, the Neuroflux team acknowledges that they are still improvable.

Aligning with the UN’s SDG 3

Neuroflux directly supports SDG 3: Good Health and Well-being by:

1. Improving early detection capabilities for one of the most aggressive brain cancers

2. Enhancing diagnostic accuracy in resource-limited settings through computational efficiency

3. Providing explainable results that build trust and adoption among clinicians

4. Creating an open-source solution accessible to low-resource healthcare systems worldwide

Get Started with Neuroflux Today

The complete codebase, along with comprehensive documentation, is now available on GitHub. The Neuroflux team extends an invitation towards researchers, clinicians, and developers to explore, use, and contribute to Neuroflux. Together, we can ethically and practically advance the field of AI in neuro-oncology to make a meaningful difference in the lives of patients.

For more information, collaboration opportunities or technical support, kindly contact us.