FlexiCT-3D

FlexiCT-3D is the whole-volume CT encoder in the FlexiCT family.

Input and preprocessing

Default processor output shape is [B, 1, 160, 160, 160].

The default preset orients/resamples path inputs to LPS at 2 mm spacing when spacing is available, clips HU to [-1000, 1000], z-score normalizes, pads with the tensor minimum to at least 160^3, and center crops to 160^3.

Optional presets:

• local_path: pad to a cube, then trilinear-resize to 160^3 for arbitrary local CT files.
• retrieval_roi: crop an ROI from coordinates, a mask, or a bounding box, pad boundary crops, then resize to 160^3.

from transformers import AutoImageProcessor, AutoModel

processor = AutoImageProcessor.from_pretrained("ricklisz123/FlexiCT-3D", trust_remote_code=True)
model = AutoModel.from_pretrained("ricklisz123/FlexiCT-3D", trust_remote_code=True)

inputs = processor("/path/to/ct.nii.gz", return_tensors="pt")
outputs = model(**inputs)
cls_token = outputs.cls_token
patch_tokens = outputs.patch_tokens

Outputs

cls_token is a global volume embedding. patch_tokens are 3D patch embeddings.

Limitations

This model is a research feature extractor. The local_path and retrieval_roi presets are convenience paths and may not match training or evaluation preprocessing exactly.