vbai 0.1.9

Visual Brain AI - Multi-task brain MRI analysis library

Vbai - Visual Brain AI

A PyTorch-based deep learning library for multi-task brain MRI analysis. Train models for dementia classification, brain tumor detection, or both simultaneously with just a few lines of code.

Features

• Flexible Task Selection: Train for dementia only, tumor only, or both simultaneously
• Easy to Use: Keras-like API for quick training
• Task-Specific Attention: Separate attention mechanisms for each task
• Visualization: Built-in attention heatmap visualization (adapts to active tasks)
• Configurable: YAML/JSON configuration support
• Production Ready: Export and deploy trained models

Supported Classifications

Dementia (6 classes):

• AD Alzheimer's Disease
• AD Mild Demented
• AD Moderate Demented
• AD Very Mild Demented
• CN Non-Demented (Cognitively Normal)
• PD Parkinson's Disease

Brain Tumor (4 classes):

• Glioma
• Meningioma
• No Tumor
• Pituitary

Installation

# Basic installation
pip install vbai

# With all optional dependencies
pip install vbai[full]

# Development installation
git clone https://github.com/Neurazum-AI-Department/vbai.git
cd vbai
pip install -e .[dev]

Quick Start

Training a Multi-Task Model (Both Dementia & Tumor)

import vbai

# Create model for both tasks (default)
model = vbai.MultiTaskBrainModel(variant='q')  # 'q' for quality, 'f' for fast

# Prepare dataset
dataset = vbai.UnifiedMRIDataset(
    dementia_path='./data/dementia/train',
    tumor_path='./data/tumor/train',
    is_training=True
)

# Create trainer
trainer = vbai.Trainer(
    model=model,
    lr=0.0005,
    device='cuda'
)

# Train
history = trainer.fit(
    train_data=dataset,
    epochs=10,
    batch_size=32
)

# Save model
trainer.save('brain_model.pt')

Training a Single-Task Model

import vbai

# Dementia only model
dementia_model = vbai.MultiTaskBrainModel(
    variant='q',
    tasks=['dementia']  # Only dementia classification
)

# Tumor only model
tumor_model = vbai.MultiTaskBrainModel(
    variant='q',
    tasks=['tumor']  # Only tumor detection
)

# Dataset for single task
dementia_dataset = vbai.UnifiedMRIDataset(
    dementia_path='./data/dementia/train',
    tumor_path=None,  # Not needed for dementia-only
    is_training=True
)

# Train as usual
trainer = vbai.Trainer(model=dementia_model, lr=0.0005)
trainer.fit(train_data=dementia_dataset, epochs=10)

Making Predictions

import vbai

# Load trained model
model = vbai.load('brain_model.pt', device='cuda')

# Single image prediction
result = model.predict('brain_scan.jpg')

# Multi-task model returns both predictions
if result.dementia_class:
    print(f"Dementia: {result.dementia_class} ({result.dementia_confidence:.1%})")
if result.tumor_class:
    print(f"Tumor: {result.tumor_class} ({result.tumor_confidence:.1%})")

# With attention visualization
result = model.predict('brain_scan.jpg', return_attention=True)
vis = vbai.VisualizationManager()
vis.visualize('brain_scan.jpg', result, save=True)
# Visualization adapts to show only active task panels

Using Callbacks

import vbai

model = vbai.MultiTaskBrainModel(variant='q')

# Setup callbacks
callbacks = [
    vbai.EarlyStopping(monitor='val_loss', patience=5),
    vbai.ModelCheckpoint(
        filepath='checkpoints/model_{epoch:02d}.pt',
        monitor='val_loss',
        save_best_only=True
    )
]

trainer = vbai.Trainer(model=model, callbacks=callbacks)
trainer.fit(train_data, val_data, epochs=50)

Configuration

import vbai

# Use preset configurations
config = vbai.get_default_config('quality')  # 'default', 'fast', 'quality', 'debug'

# Or customize
model_config = vbai.ModelConfig(
    variant='q',
    tasks=['dementia', 'tumor'],  # Choose which tasks to enable
    dropout=0.3,
    use_edge_branch=True
)

training_config = vbai.TrainingConfig(
    epochs=20,
    batch_size=16,
    lr=0.0001,
    scheduler='cosine'
)

# Save/Load configs
model_config.save('model_config.yaml')
loaded_config = vbai.ModelConfig.load('model_config.yaml')

Command Line Interface

Training

# Train both tasks
vbai-train --dementia_path ./data/dementia --tumor_path ./data/tumor --epochs 10

# Train dementia only
vbai-train --dementia_path ./data/dementia --tasks dementia --epochs 10

# Train tumor only
vbai-train --tumor_path ./data/tumor --tasks tumor --epochs 10

# Advanced options
vbai-train --dementia_path ./data/dementia --tumor_path ./data/tumor \
    --variant q --tasks dementia tumor --epochs 20 --batch_size 16 --lr 0.0001

Prediction

# Make prediction
vbai-predict --model brain_model.pt --image brain_scan.jpg

# With visualization
vbai-predict --model brain_model.pt --image brain_scan.jpg --visualize --output result.png

# JSON output
vbai-predict --model brain_model.pt --image brain_scan.jpg --json

Model Variants

Variant	Layers	Channels	Speed	Accuracy
f (fast)	3	32-64-128	Fast	Good
q (quality)	4	64-128-256-512	Slower	Better

Task Selection

Tasks Parameter	Description	Use Case
['dementia', 'tumor']	Both tasks (default)	Multi-task learning
['dementia']	Dementia only	Specialized dementia detection
['tumor']	Tumor only	Specialized tumor detection

Dataset Structure

Your dataset should be organized as follows:

data/
├── dementia/
│   ├── train/
│   │   ├── AD_Alzheimer/
│   │   ├── AD_Mild_Demented/
│   │   ├── AD_Moderate_Demented/
│   │   ├── AD_Very_Mild_Demented/
│   │   ├── CN_Non_Demented/
│   │   └── PD_Parkinson/
│   └── val/
│       └── ...
└── tumor/
    ├── train/
    │   ├── Glioma/
    │   ├── Meningioma/
    │   ├── No_Tumor/
    │   └── Pituitary/
    └── val/
        └── ...

Note: You only need the dataset for the task(s) you're training. For single-task training, only the relevant dataset directory is required.

API Reference

Core Classes

• MultiTaskBrainModel - Main model class (supports single and multi-task)
• UnifiedMRIDataset - Dataset for training (handles missing task data)
• Trainer - Training loop manager
• VisualizationManager - Attention heatmap visualization (adapts to active tasks)

Configuration

• ModelConfig - Model architecture settings (includes tasks parameter)
• TrainingConfig - Training hyperparameters
• get_default_config() - Preset configurations

Callbacks

• EarlyStopping - Stop when no improvement
• ModelCheckpoint - Save best/all checkpoints
• TensorBoardLogger - Log to TensorBoard

Examples

See the examples/ directory for complete examples:

• train_basic.py - Basic training example
• train_advanced.py - Advanced training with callbacks
• inference.py - Model inference

Citation

If you use Vbai in your research, please cite:

@software{vbai,
  title = {Vbai: Visual Brain AI Library},
  author = {Neurazum},
  year = {2025},
  url = {https://github.com/Neurazum-AI-Department/vbai}
}

License

MIT License - see LICENSE for details.

Contributing

Is being planned...

Support

• Website: Neurazum - HealFuture
• Email: [e-mail]

---

Neurazum AI Department