How miagrouput/augmentations-for-the-unknown works
This repository implements MRI image augmentations for out-of-distribution segmentation, plus nnU-Net training variants and a dataset-generation utility. A benchmark-generation workflow loads image/label pairs, applies configured transforms at severity levels, and writes transformed samples. The training branch adds augmentation and loss variants around nnU-Net. The README documents the research goal and pretrained models/testing data, but not a runtime pipeline; several trainer variants and transform implementations were not sampled, so their connections are omitted where unverified.
Main entry points: nnUNetTrainerNoDAV2FLCE_MixUp_AFA.py AFA MixUp trainer · nnUNetTrainerV2_CutMixUpAFA.py CutMix trainer · generator.py Dataset generator · defaults.py Default settings
Components 28
Dataset generation
- Dataset generator
medg/data/generator.py - Data utilities
medg/data/utils.py - Attribute utilities
medg/utils/misc.py - NIfTI writer
medg/utils/writer.py
MRI augmentations
- Default settings
medg/transforms/defaults.py - MRI artifacts
medg/transforms/additional/random_spike_n.py - K-space sampling
medg/transforms/additional/fastMRI/subsample.py - K-space transforms
medg/transforms/additional/k_space_subsampling.py - Isotropic downsampling
medg/transforms/additional/iso_downsample.py - Additional transforms
medg/transforms/additional/defaults.py
Segmentation training
- FLCE trainer
nnunet/training/network_training/nnUNetTrainerNoDAV2FLCE.py - AFA MixUp trainer
nnunet/training/network_training/nnUNetTrainerNoDAV2FLCE_MixUp_AFA.py - CutMix trainer
nnunet/training/network_training/nnUNetTrainerV2_CutMixUpAFA.py - AFA augmentation
nnunet/training/data_augmentation/afa.py - MixUp augmentation
nnunet/training/data_augmentation/mixup.py - CutMix augmentation
nnunet/training/data_augmentation/cutmix.py - Segmentation losses
nnunet/training/loss_functions/dice_loss.py - Deep supervision loss
nnunet/training/loss_functions/deep_supervision.py - Cross entropy
nnunet/training/loss_functions/crossentropy.py - Focal loss
nnunet/training/loss_functions/focal_loss.py - Generic U-Net
ml/model/segmentation/nnunet/generic_unet.py - Model helper
ml/model/segmentation/nnunet/nnunet_helper.py - DuBIN normalization
nnunet/network_architecture/custom_modules/dubin.py - Dual batch norm
nnunet/network_architecture/custom_modules/dubn.py
Model analysis
- Latent model analysis
model_surgery.py
Other components
- Input scans
- Generated dataset
- Scans and labels
Connections 25
- Input scans → Dataset generator: provides scans
medg/data/generator.py - Dataset generator → Default settings: loads defaults
medg/data/generator.py - Dataset generator → Attribute utilities: uses helpers
medg/data/generator.py - Dataset generator → NIfTI writer: writes samples
medg/data/generator.py - Default settings → MRI artifacts: configures
medg/transforms/defaults.py - Default settings → Additional transforms: configures
medg/transforms/defaults.py - Additional transforms → Isotropic downsampling: registers
medg/transforms/additional/defaults.py - Additional transforms → K-space transforms: registers
medg/transforms/additional/defaults.py - K-space transforms → K-space sampling: uses
medg/transforms/additional/k_space_subsampling.py - NIfTI writer → Generated dataset: saves NIfTI
medg/utils/writer.py - AFA MixUp trainer → FLCE trainer: extends
nnunet/training/network_training/nnUNetTrainerNoDAV2FLCE_MixUp_AFA.py - AFA MixUp trainer → AFA augmentation: applies
nnunet/training/network_training/nnUNetTrainerNoDAV2FLCE_MixUp_AFA.py - AFA MixUp trainer → MixUp augmentation: applies
nnunet/training/network_training/nnUNetTrainerNoDAV2FLCE_MixUp_AFA.py - AFA MixUp trainer → Focal loss: uses
nnunet/training/network_training/nnUNetTrainerNoDAV2FLCE_MixUp_AFA.py - CutMix trainer → AFA augmentation: applies
nnunet/training/network_training/nnUNetTrainerV2_CutMixUpAFA.py - CutMix trainer → MixUp augmentation: applies
nnunet/training/network_training/nnUNetTrainerV2_CutMixUpAFA.py - CutMix trainer → CutMix augmentation: applies
nnunet/training/network_training/nnUNetTrainerV2_CutMixUpAFA.py - CutMix trainer → Segmentation losses: uses
nnunet/training/network_training/nnUNetTrainerV2_CutMixUpAFA.py - FLCE trainer → Deep supervision loss: uses
nnunet/training/network_training/nnUNetTrainerNoDAV2FLCE.py - Model helper → Generic U-Net: constructs
ml/model/segmentation/nnunet/nnunet_helper.py - Model helper → DuBIN normalization: converts normalization
ml/model/segmentation/nnunet/nnunet_helper.py - DuBIN normalization → Dual batch norm: uses
nnunet/network_architecture/custom_modules/dubin.py - Scans and labels → Latent model analysis: analyzes scans
model_surgery.py - Latent model analysis → Model helper: loads model
model_surgery.py - Latent model analysis → Generic U-Net: runs inference
model_surgery.py