monai-seg-baseline-scaffold

name	monai-seg-baseline-scaffold
description	Generates a complete MONAI-based training scaffold for medical image segmentation or classification tasks. Creates training scripts, inference scripts, configuration files, data loading pipelines, and model definitions.
version	0.1.0
author	Medical AI Team <[e-mail]>
license	MIT
tags	["monai","segmentation","classification","training","scaffold","deep-learning","pytorch"]
maturity	scaffold
category	segmentation-ml
requires_extras	["core","ml"]

Overview

The monai-seg-baseline-scaffold skill generates a complete training scaffold for medical image analysis using the MONAI framework. It automates the creation of:

• Training script (train.py) - Full training loop with validation
• Inference script (inference.py) - Model prediction on new data
• Configuration file (config.yaml) - All hyperparameters and settings
• Data loader (data_loader.py) - MONAI-compatible data pipeline
• Model definition (model.py) - Network architecture definitions

This skill is essential for rapidly prototyping medical image analysis projects, ensuring consistent code structure and following MONAI best practices.

Installation

pip install -r requirements.txt

Or install the specific dependencies:

pip install pyyaml jinja2 monai torch numpy nibabel

Dependencies:

• Python >= 3.10
• PyYAML >= 6.0
• Jinja2 >= 3.0
• MONAI >= 1.3.0
• PyTorch >= 2.0.0
• NumPy >= 1.20.0
• NiBabel >= 5.0.0 (for medical image I/O)

Usage

python scripts/main.py --dataset-path PATH --output-dir DIR [OPTIONS]

Arguments

Argument	Required	Description
--dataset-path	Yes	Path to the dataset directory
--output-dir	Yes	Output directory for generated files

Options

Option	Default	Description
--task	segmentation	Task type: segmentation or classification
--model	unet	Model architecture: unet, swinunetr, resnet
--config-only	false	Generate only config.yaml, no scripts
--num-classes	2	Number of output classes
--input-channels	1	Number of input channels
--image-size	96	Input image size (isotropic)
--batch-size	4	Training batch size
--learning-rate	0.0001	Initial learning rate
--max-epochs	100	Maximum training epochs
--device	cuda	Training device: cuda, cpu, mps
--amp	true	Enable automatic mixed precision
--cache-rate	1.0	Cache rate for dataset (0.0-1.0)
--num-workers	4	Number of data loading workers
--verbose, -v	false	Enable verbose output
--version	-	Show version and exit
--help, -h	-	Show help message and exit

Input Schema

Input is a dataset directory plus scaffold-generation parameters. The skill generates project files and configuration templates; it does not train or validate a model.

Output Schema

{
  "success": true,
  "result": {
    "output_dir": "./experiment",
    "generated_files": ["train.py", "inference.py", "config.yaml"],
    "task": "segmentation",
    "model": "unet"
  },
  "metadata": {
    "timestamp": "2024-01-15T10:30:00Z",
    "duration_ms": 250,
    "version": "0.1.0"
  }
}

Exit Codes

Code	Meaning
0	Success - all files generated
1	Validation error - invalid arguments or config
2	Runtime error - file I/O or template error
3	Usage error - missing required arguments

Examples

Basic usage (segmentation with default UNet)

python scripts/main.py \
  --dataset-path /data/brain_tumor \
  --output-dir ./brain_tumor_experiment

Classification task with ResNet

python scripts/main.py \
  --dataset-path /data/chest_xray \
  --output-dir ./chest_xray_exp \
  --task classification \
  --model resnet \
  --num-classes 3

Generate config only

python scripts/main.py \
  --dataset-path /data/experiment \
  --output-dir ./config_only \
  --config-only

Model Architectures

UNet (Default)

Standard U-Net architecture with configurable channels and strides.

SwinUNETR

Vision Transformer-based architecture for 3D medical image segmentation.

ResNet (Classification)

Residual network for classification tasks.

References

• MONAI Documentation
• MONAI GitHub
• PyTorch Documentation

Disclaimer

For Research Use Only - This scaffold generates starter code and requires human review before any real-world or clinical use. It is not production-ready.

Changelog

0.1.0 (2026-03-15)

• Initial release
• Support for segmentation and classification tasks
• Model architectures: UNet, SwinUNETR, ResNet
• Jinja2 template-based code generation
• Complete training loop with validation