NeuroGenAI: A Full-Stack Deep Learning-Based Brain Tumor Detection System with AI-Assisted Clinical Report Generation

Abstract

Every minute separating a suspicious MRI from a physician-readable report is a minute of clinical uncertainty—and in neuro-oncology, uncertainty compounds rapidly into harm. NeuroGenAI confronts this reality by fusing a VGG16 convolutional classifier with a large language model reporting engine to produce, within a single browser interaction, both a tumour classification and a structured clinical narrative ready for physician review. The classifier operates across four neuro-oncological categories (Glioma, Meningioma, Pituitary Tumour, No Tumour) on gadolinium-free MRI input; the Google Gemini language model then converts the raw prediction into section-organised prose (Findings, Impression, Recommendations) conditioned on patient metadata. Surrounding these inference components is a production-grade web stack: React.js for the clinician-facing interface, Django REST Framework for backend orchestration, PostgreSQL for relational patient records, and Cloudinary for cloud-resident media. On a 1,400-image holdout partition of the Br35H dataset, the system posts a weighted-average precision of 0.96, a Pituitary-class recall of 0.99, and an end-to-end scan-to-report latency of roughly 2.8 seconds—figures that together make a practical, rather than aspirational, case for AI-assisted neuro-oncology at the point of care.

Introduction

Brain tumors often present symptoms such as headaches, seizures, and cognitive changes that resemble other neurological conditions, making MRI scans essential for diagnosis. However, interpreting MRI images is a complex and time-consuming task that depends heavily on radiologists, whose performance may be affected by workload, experience, and resource availability. While deep learning models have achieved high accuracy in classifying brain tumors from MRI scans, most existing systems only provide numerical predictions without generating the structured clinical reports required for real-world medical decision-making.

To address this gap, the proposed system, NeuroGenAI, combines artificial intelligence-based tumor classification with automated clinical report generation. The system is designed as a complete clinical platform that not only identifies tumors but also produces physician-ready documentation and supports secure deployment in healthcare environments.

Key Contributions of NeuroGenAI

1. Accurate Tumor Classification
   • Uses a custom-trained VGG16 Convolutional Neural Network (CNN).
   • Trained on the Br35H Brain Tumor Dataset containing 7,023 MRI images.
   • Classifies MRI scans into four categories:
     • Glioma
     • Meningioma
     • Pituitary Tumor
     • No Tumor
   • Achieves approximately 96.5% weighted average accuracy.
2. Automated Clinical Report Generation
   • Integrates a Google Gemini Large Language Model (LLM).
   • Converts classification results and patient information into structured medical reports.
   • Automatically generates:
     • Findings
     • Medical Impression
     • Recommended Next Steps
   • Reduces the documentation burden on clinicians.
3. Deployment-Ready Healthcare Platform
   • Built using:
     • React.js (Frontend)
     • Django REST Framework (Backend)
     • PostgreSQL (Database)
     • Cloudinary (Cloud Storage)
   • Supports role-based access control, secure authentication using JWT, and cloud-based media management.

Literature Review Findings

Research shows a progression from traditional machine learning methods using handcrafted features to deep learning models such as VGG16, ResNet, and Inception, which learn features directly from MRI images and achieve over 95% classification accuracy. However, existing systems generally suffer from two limitations:

• Lack of interpretability and clinical report generation.
• Absence of deployment infrastructure required for real hospital use.

NeuroGenAI addresses both challenges by integrating image classification and automated report creation within a single platform.

Problem Statement

Current AI systems have largely solved the accuracy problem in tumor classification but have not solved the workflow problem. A prediction alone cannot be directly used in patient records, tumor board discussions, or treatment planning. Hospitals require structured reports, secure storage, authenticated access, and integrated workflows. NeuroGenAI is designed to bridge this gap.

System Architecture

The platform follows a three-tier Service-Oriented Architecture (SOA):

1. Presentation Layer (Frontend)

• Technician portal for MRI upload and patient data entry.
• Physician dashboard for viewing results, scan history, and downloading reports.
• Role-based access control and secure authentication.

2. Application and Intelligence Layer (Backend)

• MRI preprocessing and classification using VGG16.
• LLM-powered report generation through Google Gemini.
• PDF report creation and secure API communication.

3. Data Persistence Layer

• PostgreSQL stores patient and scan records.
• Cloudinary stores MRI images and generated reports.
• Maintains secure links between medical records and reports.

Methodology

The development process involved three stages:

1. Dataset Preparation
   • MRI images resized to 224×224 pixels.
   • Pixel normalization applied.
   • Data augmentation through rotation, flipping, and zooming to improve model generalization.
2. Model Training
   • VGG16 architecture selected due to its effective feature extraction capability.
   • Uses stacked 3×3 convolution filters to learn increasingly complex tumor characteristics.
3. Report Generation Integration
   • AI predictions and patient metadata are sent to Gemini.
   • Structured reports are automatically generated and converted into downloadable PDFs.

Conclusion

NeuroGenAI represents a quantum leap in clinical neuroimaging by integrating deep learning classification with generative AI narrative synthesis. The system effectively overcomes the “black box” limitations of traditional AI models, providing radiologists with accurate multi-class tumor detection paired with structured medical rationale. Measured against the two-register problem articulated in Section III, the platform delivers on both fronts: VGG16 achieves 96.5% classification accuracy with a Pituitary-class recall of 0.99, addressing the diagnostic accuracy requirement; the Gemini reporting engine converts numerical predictions into physician-ready documentation in under 1.2 seconds, addressing the workflow integration requirement. The React–Django–PostgreSQL–Cloudinary production stack provides the surrounding security and scalability infrastructure that elevates the system from a research prototype to a clinically deployable product. Three directions present themselves as natural extensions. First, saliency map overlays (Grad-CAM or integrated gradients) could annotate the MRI image with the spatial regions that drove the classification decision, adding a visual dimension to the textual report and enabling radiologist calibration of the AI’s attentional strategy. Second, federated learning protocols would allow multi-institutional training without centralising patient data, expanding the diversity of tumour morphologies seen during optimisation. Third, fusion of imaging predictions with structured clinical data (symptom duration, biomarker profiles) could shift the system from single-modality classification toward genuinely multi-factorial diagnostic support—the direction that clinical AI must ultimately travel to match the integrative reasoning that experienced clinicians perform.

References

[1] K. Lata et al., “Deep Learning-Based Brain Tumor Detection in Privacy-Preserving Smart Health Care Systems,” IEEE Access, 2024. [2] Q. Yao et al., “Accurate Detection of Brain Tumor Lesions From Medical Images Based on Improved YOLOv8 Algorithm,” IEEE Access, 2024. [3] S. Solanki et al., “Brain Tumor Detection and Classification Using Intelligence Techniques: An Overview,” IEEE Access, 2023. [4] K. N. Rao et al., “An Efficient Brain Tumor Detection and Classification Using Pre-trained CNN Models,” Heliyon, 2024. [5] S. Pereira et al., “Brain Tumor Segmentation Using CNNs in MRI Images,” IEEE Transactions on Medical Imaging, 2016. [6] A. Rehman et al., “Deep Learning-Based Framework for Brain Tumors Classification,” Circuits, Systems, and Signal Processing, 2020. [7] K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale Image Recognition,” arXiv:1409.1556, 2014. [8] Gemini Team, “Gemini: A Family of Highly Capable Multimodal Models,” arXiv:2312.11805, 2023. [9] A. Hamada, “Br35H Brain Tumor Dataset,” Kaggle, 2020.

Copyright

Copyright © 2026 Tushar Nangare, Kalpesh Patil, Sarvesh Namra, Neha Patil, Prof. Digvijay Patil. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.