medicai 0.0.4

Keras-based medical 2D and 3D image analysis library.

Medic-AI is a Keras based library designed for medical 2D and 3D image analysis using machine learning techniques. Its core strengths include:

• Backend Agnostic: Compatible with tensorflow, torch, and jax.
• User-Friendly API: High-level interface for transformations and model creation of both 2D and 3D.
• Scalable Execution: Supports training and inference on single/multi-GPU and TPU-VM setups.
• Essential Components: Includes standard medical specific metrics and losses, such as Dice. Support GradCAM for segmentation and classification on both 2D and 3D input.
• Optimized 3D Inference: Offers an efficient sliding-window method and callback for volumetric data.

📋 Table of Contents

1. Installation
2. Features
3. Documentation
4. Acknowledgements
5. Citation

🛠 Installation

PyPI version:

pip install medicai

This installs medicai and keras, but leaves backend runtime selection to you. Install tensorflow, torch, or jax separately based on your workflow.

Installing from source GitHub: (recommended)

pip install git+https://github.com/innat/medic-ai.git

Using uv for local development:

uv venv --python 3.10
source .venv/bin/activate
uv pip install -e .[dev]

Optional extras:

pip install "medicai[docs]"
pip install "medicai[test]"
pip install "medicai[dev]"

Quick Overview

For details end-to-end training workflow, please check the guide section.

from medicai.models import SwinUNETR, UNet
from medicai.models import SwinTiny, SwinTinyV2
from medicai.models import SwinBackbone, SwinBackboneV2

# Build 3D model.
model = SwinUNETR(
    encoder_name='swin_tiny_v2', input_shape=(96,96,96,1)
)
model = UNet(
    encoder_name='densenet121', input_shape=(96,96,96,1)
)

# Build 2D model.
model = SwinUNETR(
    encoder_name='swin_tiny_v2', input_shape=(96,96,1)
)
model = UNet(
    encoder_name='densenet121', input_shape=(96,96,1)
)

# Build with pre-built encoder.
encoder = SwinTiny(
    input_shape=(96,96,96,1),
    patch_size=2, 
    downsampling_strategy='swin_unetr_like'
)
model = SwinUNETR(encoder=encoder)

# Build with custom encoder.
custom_encoder = SwinBackboneV2(
    input_shape=(64, 128, 128, 1),
    embed_dim=48,
    window_size=8,
    patch_size=2,
    downsampling_strategy='swin_unetr_like'
)
model = SwinUNETR(encoder=custom_encoder)

📊 Features

Available Models : The following table lists the currently supported models along with their supported input modalities, primary tasks, and underlying architecture type. The model inputs can be either 3D (depth × height × width × channel) or 2D (height × width × channel).

Model	Supported Modalities	Primary Task	Architecture Type
DenseNet	2D, 3D	Classification	CNN
ResNet-V1,V2	2D, 3D	Classification	CNN
ResNeXt	2D, 3D	Classification	CNN
SE-ResNet	2D, 3D	Classification	CNN
SE-ResNeXt	2D, 3D	Classification	CNN
Xception	2D, 3D	Classification	CNN
EfficientNet-V1,V2	2D, 3D	Classification	CNN
ConvNeXt-V1,V2	2D, 3D	Classification	CNN
ViT	2D, 3D	Classification	Transformer
MiT	2D, 3D	Classification	Transformer
Swin Transformer-V1,V2	2D, 3D	Classification	Transformer
UNet	2D, 3D	Segmentation	CNN
UNet++	2D, 3D	Segmentation	CNN
AttentionUNet	2D, 3D	Segmentation	CNN
DeepLabV3Plus	2D, 3D	Segmentation	CNN
UPerNet	2D, 3D	Segmentation	CNN
UNETR	2D, 3D	Segmentation	Transformer
UNETR++	2D, 3D	Segmentation	Transformer
SwinUNETR	2D, 3D	Segmentation	Transformer
SwinUNETR-V2	2D, 3D	Segmentation	Transformer
TransUNet	2D, 3D	Segmentation	Transformer
SegFormer	2D, 3D	Segmentation	Transformer

Available Transformation: The following preprocessing and transformation methods are supported for volumetric data. The following layers are implemented with TensorFlow operations. It can be used in the tf.data API or a Python data generator and is fully compatible with multiple backends, tf, torch, jax in training and inference, supporting both GPUs and TPUs.

CropForeground
NormalizeIntensity
Orientation
RandomCropByPosNegLabel
RandomFlip
RandomRotate90
RandomRotate
RandomCutOut
RandomShiftIntensity
RandomSpatialCrop
Resize
ScaleIntensityRange
Spacing

📚 Documentation

To learn more about models, transformations, and training, please visit the Read the Docs documentation: medicai.readthedocs.io

🤝 Contributing

Please check the contribution guide here.

🙏 Acknowledgements

This project is greatly inspired by MONAI.

📝 Citation

If you use medicai in your research or educational purposes, please cite it using the metadata from our CITATION.cff file.