Segmentation models

Here the four best onnx models are published as described in the thesis.

All models expect the input as [batch_size, 3, 320, 896] Where 3 is in the sagittal plane, 320 is in the coronal plane and 896 is in the transverse plane. If your image is smaller, then put the image at the "bottom" in the transverse plane and pad with 0 at the top until you have the required shape.

• 2023-11-06_Seg2_Unet_resnet152_896px.onnx is a 2-class (+background) segmentation model. It separates the image into background (0), vertebrae (1) and IVDs (2), therefore outputting [batch_size, 320, 896].
• 2023-11-09_Seg25_Unet_resnet152_vert-only_896px_no-input-seg.onnx is a 25-class (+background) "instance" segmentation model. It separates the image into background (0) and vertebrae (1, 2, ..., 25), therefore outputting [batch_size, 320, 896].

Naming explained for the models:

Semantic segments in the name are split by _, and grouped by -.

• 2023-11-06: train date
• Seg<n>: Segmentation model with n output classes. Generally, a 2-class model outputs vertebrae and IVDs, a 24-class model outputs each IVD separately, a 25-class model outputs each vertebra separately, and a 49-class model outputs each vertebra and IVD separately.
• Unet: model architecture
• resnet152: model backbone
• vert-only: vertebrae only (implies 1-class or 25-class)
• 896px: expected input height in transverse plane
• no-input-seg: model does not expect 2-class segmentation as input.