wdika/SEG_UNet_ISLES2022SubAcuteStroke

Model Overview

AttentionUNet for MRI Segmentation on the ISLES2022SubAcuteStroke dataset.

ATOMMIC: Training

To train, fine-tune, or test the model you will need to install ATOMMIC. We recommend you install it after you've installed latest Pytorch version.

pip install atommic['all']

How to Use this Model

The model is available for use in ATOMMIC, and can be used as a pre-trained checkpoint for inference or for fine-tuning on another dataset.

Corresponding configuration YAML files can be found here.

Automatically instantiate the model

pretrained: true
checkpoint: https://huggingface.co/wdika/SEG_UNet_ISLES2022SubAcuteStroke/blob/main/SEG_UNet_ISLES2022SubAcuteStroke.atommic
mode: test

Usage

You need to download the ISLES 2022 Sub Acute Stroke dataset to effectively use this model. Check the ISLES2022SubAcuteStroke page for more information.

Model Architecture

model:
  model_name: SEGMENTATIONUNET
  segmentation_module: UNet
  segmentation_module_input_channels: 3
  segmentation_module_output_channels: 1
  segmentation_module_channels: 32
  segmentation_module_pooling_layers: 5
  segmentation_module_dropout: 0.0
  segmentation_module_normalize: false
  segmentation_loss:
    dice: 1.0
  dice_loss_include_background: true  # always set to true if the background is removed
  dice_loss_to_onehot_y: false
  dice_loss_sigmoid: false
  dice_loss_softmax: false
  dice_loss_other_act: none
  dice_loss_squared_pred: false
  dice_loss_jaccard: false
  dice_loss_flatten: false
  dice_loss_reduction: mean_batch
  dice_loss_smooth_nr: 1e-5
  dice_loss_smooth_dr: 1e-5
  dice_loss_batch: true
  dice_metric_include_background: true  # always set to true if the background is removed
  dice_metric_to_onehot_y: false
  dice_metric_sigmoid: false
  dice_metric_softmax: false
  dice_metric_other_act: none
  dice_metric_squared_pred: false
  dice_metric_jaccard: false
  dice_metric_flatten: false
  dice_metric_reduction: mean_batch
  dice_metric_smooth_nr: 1e-5
  dice_metric_smooth_dr: 1e-5
  dice_metric_batch: true
  segmentation_classes_thresholds: [ 0.5 ]
  segmentation_activation: sigmoid
  magnitude_input: true
  log_multiple_modalities: true  # log all modalities in the same image, e.g. T1, T2, T1ce, FLAIR will be concatenated
  normalization_type: minmax
  normalize_segmentation_output: true
  complex_data: false

Training

  optim:
    name: adamw
    lr: 1e-4
    betas:
      - 0.9
      - 0.999
    weight_decay: 0.0
    sched:
        name: CosineAnnealing
        min_lr: 0.0
        last_epoch: -1
        warmup_ratio: 0.1

trainer:
  strategy: ddp_find_unused_parameters_false
  accelerator: gpu
  devices: 1
  num_nodes: 1
  max_epochs: 50
  precision: 16-mixed  # '16-mixed', 'bf16-mixed', '32-true', '64-true', '64', '32', '16', 'bf16'
  enable_checkpointing: false
  logger: false
  log_every_n_steps: 50
  check_val_every_n_epoch: -1
  max_steps: -1

Performance

Evaluation can be performed using the segmentation evaluation script for the segmentation task, with --evaluation_type per_slice.

Results

Evaluation

ALD = 0.9088 +/- 3.953 AVD = 0.5439 +/- 3.921 DICE = 0.6946 +/- 0.5589 L-F1 = 0.7859 +/- 0.5848

Limitations

This model was trained on the ISLES2022SubAcuteStroke dataset with stacked ADC, DWI, FLAIR images and might differ in performance compared to the leaderboard results.

References

[1] ATOMMIC

[2] Petzsche MRH, Rosa E de la, Hanning U, et al. ISLES 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. Scientific Data 1 2022;9

wdika
/

SEG_UNet_ISLES2022SubAcuteStroke