dafne-monai-inference

Lightweight inference and incremental learning package for Dafne models, built on MONAI and PyTorch.

Overview

This package provides:

• Inference — sliding-window (3D) and slice-by-slice (2D) segmentation inference using trained Dafne models
• Incremental learning — online fine-tuning of a deployed model on new annotated data, with optional EWC (Elastic Weight Consolidation) regularization to prevent catastrophic forgetting
• Network architectures — DafneUnetModel and DafneDynUnetModel, wrapping MONAI's UNet and DynUNet with Dafne-compatible interfaces
• Transforms — preprocessing pipeline with anisotropy handling, spacing normalization, and bounding-box cropping

Installation

pip install dafne-monai-inference

Requirements

• Python >= 3.8
• torch >= 2.0
• monai >= 1.3
• numpy >= 1.24
• scikit-image >= 0.19

Usage

Inference

from dafne_inference.inference import run_inference

# model_obj: a DynamicTorchModel loaded from a .model file
# data_dict must contain 'image' (numpy array) and either 'affine' or 'resolution'
result = run_inference(model_obj, data_dict)
# result: dict { label_name: binary_mask (np.int8) }

Incremental learning

from dafne_inference.incremental import run_incremental_learning

# trainingData: dict with key 'image_list' (list of numpy arrays) and 'resolution'
# trainingOutputs: dict of masks corresponding to each image
run_incremental_learning(model_obj, trainingData, trainingOutputs, bs=1, minTrainImages=2)

Network architectures

from dafne_inference.networks import DafneUnetModel, DafneDynUnetModel

# Standard UNet
model = DafneUnetModel(
    spatial_dims=3,
    in_channels=1,
    out_channels=5,
    start_channel=32,
    n_levels=5
)

# Dynamic UNet (nnU-Net style)
model = DafneDynUnetModel(
    spatial_dims=3,
    in_channels=1,
    out_channels=5,
    kernels=[[3,3,3], [3,3,3], [3,3,3]],
    strides=[[1,1,1], [2,2,2], [2,2,2]]
)

License

GNU General Public License v3 or later (GPLv3+)