brain-tumor-detection

Model Card: Brain Tumor Multi-Class Classification

Model Details

Model Description

A fine-tuned ResNet18 convolutional neural network for classifying brain MRI scans into four categories of brain tumors. The model uses transfer learning with ImageNet pre-trained weights and has been adapted for medical image classification.

• Developed by: Thisen Ekanayake
• Model type: Convolutional Neural Network (ResNet18)
• Language(s): Python (PyTorch)
• License: MIT License
• Demo: https://brainet.thisenekanayake.me
• Parent Model: ResNet18 (pretrained on ImageNet)
• Model Variants: This model card focuses on the multi-class classification model (4 classes: glioma, meningioma, no tumor, pituitary). A binary classification model (tumor vs. no tumor) is also available in the GitHub repository with complete training and evaluation scripts.

Model Architecture

• Base Architecture: ResNet18
• Input: RGB images resized to 224x224 pixels
• Output: 4-class classification (glioma, meningioma, notumor, pituitary)
• Modifications:
  • Final fully connected layer replaced with 4-class output
  • Early layers frozen during training
  • Only layer4 and final FC layer fine-tuned
• Parameters: ~11M total parameters

Intended Uses

Primary Use Case

This model is designed for research and educational purposes only to demonstrate deep learning applications in medical image analysis. It can be used to:

• Study transfer learning techniques in medical imaging
• Explore brain tumor classification methodologies
• Educational demonstrations of CNN architectures
• Research prototyping and benchmarking

Out-of-Scope Uses

• NOT for clinical diagnosis or treatment decisions
• NOT for patient care or medical decision-making
• NOT a replacement for professional medical evaluation
• NOT validated for deployment in healthcare settings
• Should not be used without expert medical supervision

Training Data

Dataset Information

• Source: Brain Tumor Segmentation(BraTS2020)
• Dataset Link: https://www.kaggle.com/datasets/awsaf49/brats2020-training-data
• Classes:
  1. Glioma
  2. Meningioma
  3. No Tumor
  4. Pituitary Tumor

Dataset Statistics

• Training samples: 5712
• Testing samples: 1311
• Class Distribution (Training):
  • Imbalanced dataset with varying samples per class
  • Weighted sampling used to address class imbalance
  • Class weights applied during training

Data Preprocessing

• Image Resizing: 224x224 pixels
• Normalization: ImageNet statistics (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
• Training Augmentations:
  • Random horizontal flip (p=0.5)
  • Random rotation (±10 degrees)
  • Color jitter (brightness=0.2, contrast=0.2)
• Test Preprocessing: Resize and normalize only (no augmentation)

Known Dataset Limitations

• Limited diversity in imaging protocols and scanner types
• Potential demographic biases in patient populations
• Images may not represent all tumor subtypes or presentations
• Dataset size may limit generalization to rare cases
• Quality and annotation consistency may vary

Model Performance

Overall Metrics (Test Set)

Evaluated on 1,311 test samples with 1,298 correct predictions (13 errors):

Per-Class Performance

Detailed Analysis by Class

Glioma:

• Correctly classified: 289/300 (96.33%)
• Misclassified as Meningioma: 10 cases
• Misclassified as Pituitary: 1 case
• High precision (99.66%) indicates few false positives

Meningioma:

• Correctly classified: 306/306 (100% recall)
• Perfect recall with no false negatives
• Precision of 96.53% due to 11 false positives from other classes

No Tumor:

• Perfect classification (100% precision, recall, and F1)
• 405/405 correctly identified
• No confusion with tumor classes

Pituitary:

• Correctly classified: 298/300 (99.33%)
• Misclassified as Glioma: 1 case
• Misclassified as Meningioma: 1 case
• Near-perfect performance across all metrics

Confusion Analysis

• Total Errors: 13 out of 1,311 samples (0.99% error rate)
• Most Common Error: Glioma misclassified as Meningioma (10 cases)
• Best Performance: No Tumor class (perfect classification)
• Clinical Significance: The model never confuses tumor cases with "No Tumor", which is critical for medical screening applications

Training Procedure

Training Configuration

• Framework: PyTorch
• Optimizer: Adam
• Learning Rate: 1e-4
• Batch Size: 16
• Loss Function: CrossEntropyLoss with class weights
• Device: CUDA (GPU) Nvidia GeForce RTX 4060 8GB
• Training Strategy: Transfer learning with partial fine-tuning

Training Details

1. Layer Freezing: All layers frozen except layer4 and final FC layer
2. Class Imbalance Handling:
   • Weighted random sampling during training
   • Class-weighted loss function
3. Validation: Evaluated on test set after each epoch
4. Early Stopping: Not implemented (trained for full 10 epochs)

Computational Requirements

• Training Time: Hardware dependent
• GPU Memory: Suitable for single GPU training
• Inference Time: Fast inference suitable for real-time applications

Limitations and Biases

Technical Limitations

1. Input Constraints:

   • Requires specific image dimensions (224x224)
   • Expects RGB format with ImageNet normalization
   • May not handle varying image qualities well

2. Performance Limitations:

   • Trained on limited dataset size
   • May not generalize to images from different scanners or protocols
   • Performance on edge cases and rare tumor types unknown

3. Architectural Limitations:

   • ResNet18 is relatively shallow; deeper models might capture more complex patterns
   • Single-view classification (doesn't utilize 3D MRI volumes)
   • No attention mechanisms or interpretability features

Potential Biases

1. Dataset Bias:

   • Training data may not represent global population diversity
   • Potential geographical, demographic, or institutional biases
   • Scanner and imaging protocol biases

2. Class Imbalance:

   • Despite weighted sampling, model may still favor majority classes
   • Rare tumor presentations underrepresented

3. Annotation Bias:

   • Dependent on original dataset annotation quality
   • May inherit labeling errors or inconsistencies

Medical and Ethical Considerations

1. NOT Clinically Validated:

   • No regulatory approval (FDA, CE, etc.)
   • Not tested in real clinical workflows
   • Performance on real-world clinical data unknown

2. Risk of Misuse:

   • Should never replace professional medical judgment
   • False positives/negatives could lead to inappropriate actions if misused
   • Requires proper medical context for interpretation

3. Privacy Considerations:

   • Ensure patient data privacy when using model
   • Comply with HIPAA, GDPR, and local regulations
   • Anonymize any input images appropriately

Recommendations

For Researchers and Developers

• Use as baseline or comparison model for brain tumor classification research
• Experiment with different architectures, hyperparameters, or training strategies
• Combine with other models or modalities for improved performance
• Conduct thorough validation on your specific use case

For Educators

• Excellent demonstration of transfer learning in medical imaging
• Shows importance of handling class imbalance
• Good example of CNN fine-tuning strategies
• Can be used to teach medical AI ethics and limitations

Best Practices

1. Always validate model predictions with ground truth when available
2. Use ensemble methods or multiple models for critical applications
3. Implement uncertainty quantification to identify low-confidence predictions
4. Continuously monitor model performance on new data
5. Maintain human oversight for all predictions

How to Use

Loading the Model

import torch
import torch.nn as nn
from torchvision import models

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Load model architecture
model = models.resnet18(pretrained=False)
num_ftrs = model.fc.in_features
model.fc = nn.Linear(num_ftrs, 4)
model = model.to(device)

# Load trained weights
model.load_state_dict(torch.load("multi_class_resnet.pth"))
model.eval()

Image Preprocessing

from torchvision import transforms

test_transforms = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])

Making Predictions

with torch.no_grad():
    image = test_transforms(pil_image).unsqueeze(0).to(device)
    output = model(image)
    _, prediction = torch.max(output, 1)
    
class_names = ['glioma', 'meningioma', 'notumor', 'pituitary']
predicted_class = class_names[prediction.item()]

Citation

If you use this model in your research, please cite:

@misc{brain_tumor_classifier_2025,
  title={BRAINet - Brain Tumor Multi-Class Classification using ResNet18},
  author={Thisen Ekanayake},
  year={2025},
  url={https://brainet.thisenekanayake.me},
  note={Educational and research purposes only}
}

Dataset Citation

@misc{brats20_dataset,
  author={Awsaf},
  title={Brain Tumor Segmentation(BraTS2020)},
  year={2020},
  url={https://www.kaggle.com/datasets/awsaf49/brats2020-training-data}
}

Contact and Support

• Demo Application: https://brainet.thisenekanayake.me
• GitHub Repository: https://github.com/Thisen-Ekanayake/BRAINet
• Binary Classification Model: Available in the repository with complete training and evaluation scripts
• Issues: Submit issues via the GitHub repository
• Contributions: Welcome via pull requests to the BRAINet GitHub repository

Model Card Version

• Version: 1.0
• Date: February 2026
• Last Updated: February 2026

Glossary

• Glioma: A type of tumor that occurs in the brain and spinal cord
• Meningioma: A tumor that arises from the meninges (membranes surrounding the brain and spinal cord)
• Pituitary Tumor: A growth in the pituitary gland
• Transfer Learning: Using a pre-trained model and adapting it for a new task
• Fine-tuning: Training select layers of a pre-trained model on new data
• Class Imbalance: When training data has unequal numbers of samples per class

---

⚠️ CRITICAL DISCLAIMER

THIS MODEL IS FOR RESEARCH AND EDUCATIONAL PURPOSES ONLY

This model has NOT been validated for clinical use and should NEVER be used for:

• Medical diagnosis
• Treatment planning
• Patient care decisions
• Clinical decision support

Always consult qualified healthcare professionals for medical advice, diagnosis, and treatment. The developers assume no liability for misuse of this model in clinical or medical settings.

---

This model card follows the framework proposed by Mitchell et al. (2019) and adapted for medical AI applications.