Deep Learning-Based Brain Tumor Classification from MRI: An AI-Assisted Diagnostic Approach

🎯 Project Overview

This project aims to develop a robust, interpretable, and highly accurate AI-based diagnostic tool for classifying brain tumors from MRI images. Leveraging both custom convolutional neural networks (CNNs) and advanced transfer learning architectures such as ResNet-50, EfficientNetB0, and DenseNet-121, the system classifies MRI scans into four tumor categories: glioma, meningioma, pituitary, and no tumor.

I completed this project when taking the Teesside Deep Learning module.

---

📌 Aims and Motivation

• Reduce diagnostic errors and support faster, automated detection of brain tumors.
• Promote clinical trust through explainability (Grad-CAM) and performance stability.
• Bridge healthcare gaps in underserved regions by supporting scalable AI solutions.
• Address social, ethical, and legal concerns of AI in medicine, including:
  • Misinterpretation of AI outputs by clinicians and legal accountability.
  • Overreliance on AI leading to erosion of critical medical thinking.
  • Dataset bias affecting minority populations (e.g., ethnicity, gender).
  • Inequality in AI access across countries and healthcare systems.
  • Data privacy concerns from fast-paced AI deployment.

---

🧠 Models Implemented

1. Custom CNN Model

• Built from scratch using convolutional, pooling, and dropout layers.
• Activation functions tuned (Swish, LeakyReLU) for better convergence.
• Baseline performance established before transfer learning implementation.

2. Transfer Learning Models

• ResNet-50, EfficientNetB0, DenseNet-121
• Pre-trained on ImageNet; top layers fine-tuned for binary/multiclass tumor classification.
• Fine-tuning strategies:
  • Layer unfreezing and re-training higher layers.
  • Learning rate optimization (1e-4 to 1e-6).
  • Increased batch size (32 → 64).
  • Dropout and EarlyStopping to reduce overfitting.

3. Evaluation and Interpretability

• Metrics used: Accuracy, Precision, Recall, F1-Score
• Model explainability via Grad-CAM heatmaps to localize tumor regions.
• Frameworks: TensorFlow and Keras

---

🧪 Dataset and Preprocessing

• 📂 Source: Kaggle Brain Tumor MRI Dataset
• 🖼️ 7,022 pre-labeled MRI images across 4 classes: glioma, meningioma, pituitary, no tumor.
• Images resized to 224x224x3, padded for aspect ratio preservation.
• Additional 20-pixel padding applied to reduce corner feature loss (Chaudhary, 2022).
• Augmented using ImageDataGenerator with minor zoom, flip, and rotation for realism.

---

📊 Results & Findings

Model	Accuracy	Precision	Recall	F1-Score	Notes
EfficientNetB0	99.05%	98.99%	99.00%	99.00%	Best overall performance
DenseNet-121	99.05%	98.80%	98.95%	98.85%	Highest validation accuracy, slower
ResNet-50	98.86%	98.80%	98.85%	98.82%	Consistently strong, moderate time
Custom CNN	98.01%	97.90%	97.85%	97.87%	Baseline model after fine-tuning

---

📈 Visualizations

• Training vs. Validation Curves (Accuracy & Loss)
• Grad-CAM Heatmaps for tumor localization
• Confusion Matrices for each model

Example:

---

🔬 Ethical, Social & Legal Considerations

Risk Area	Description	Mitigation Strategy
AI Misinterpretation	Potential misdiagnoses if AI results are used without context	Transparency, human-in-the-loop review
Overreliance on AI	Erosion of clinical judgment	Retain standard medical practices, retraining programs
Legal Liability	Ambiguity in error responsibility (developers, hospitals, clinicians)	Define clear legal frameworks
Dataset Bias	Bias against minority groups (gender, ethnicity, symptoms)	Balanced data curation and evaluation
Inequitable Access	AI benefits not equally distributed across global healthcare systems	Open-source deployment, lightweight model options
Privacy & Legislation Gaps	Weak laws around data collection and AI usage	Policy advocacy and regulation development

---

📚 References

• Vimala, B. et al., (2023). Nature Scientific Reports
• Nalepa, J. et al., (2019). Frontiers in Computational Neuroscience, 13.
• Chaudhary, V. (2022). What is Padding in Neural Network?, GeeksforGeeks.
• Sheliemina, N. (2024). Health Economics and Management Review, 5(2).
• Kaggle Notebook

---

🚀 Future Work

• Expand to multi-class tumor detection using clinical subtypes.
• Integrate 3D MRI data for richer spatial analysis.
• Deploy in a real-time clinical diagnostic application.
• Explore federated learning to preserve privacy during model training.

---

👨‍💻 Author

Ekow Mills-Nketsiah
Deep Learning Researcher 📧 millsekow24@yahoo.com
🔗 linkedin Profile

---