AI4MIRO/SwinUNETR_body_segmentation

SwinUNETR

Trained by Margerie Huet Dastarac .
Training date: November 2023 .

1. Task Description

Segmentation of the body on the CT scan on a dataset of 60 oropharyngeal patients. This model can be used to clean CT scans by setting voxels value outside of the body contour to air, a typical preprocessing step for other networks.

2. Model

2.1. Architecture

Figure 1: SwinUNETR architecture

2.2. Input

• CT

2.3. Output

• BODY

2.4 Training details

• Number of epoch: 300
• Loss function: Dice loss
• Optimizer: Adam
• Learning Rate: 3e-4
• Dropout: No
• Patch size in voxels: (128,128,128)
• Data augmentation used:
  • RandSpatialCropd
  • RandFlipd axis:0
  • RandFlipd axis:1
  • RandFlipd axis:2
  • NormalizeIntensityd
  • RandScaleIntensityd factors:0.1 prob:1.0
  • RandShiftIntensityd, offsets:0.1, prob:1.0

3. Dataset

• Location: Head and neck, oropharynx
• Training set size: 60
• Resolution in mm: 3x3x3

Performance

• TBD