Gallery¶

Every figure below was made by the scripts in examples/ and is committed to the repository. Click a figure to enlarge it. How each was made and what to look for: Examples explained.

ResNet-50 looks at a cat¶

ImageNet weights, the scikit-image "chelsea" photograph (CC0).

ResNet-50, cinematic

ResNet-50, technical style ResNet-50, story style ResNet-50, cinematic 16:9 ResNet-50, cinematic light theme

neural-flow demo cat                                             # or: python examples/resnet.py
neural-flow render resnet50 -i photo.jpg --style technical -o flow.png

Vision transformer (ViT-B/16)¶

ViT-B/16, cinematic

ViT-B/16, technical

neural-flow render vit -i photo.jpg --style cinematic -o vit.png   # or: python examples/vit.py

Chest X-ray (TorchXRayVision DenseNet-121)¶

A public NIH ChestX-ray14 radiograph. 18 pathology outputs, shown as independent probabilities.

Chest X-ray, cinematic

Chest X-ray, light theme

neural-flow demo cxr                                              # or: python examples/chest_xray.py

2-D U-Net¶

Trained on synthetic microscopy images. Skip connections are drawn as bridges.

2-D U-Net, cinematic

2-D U-Net, technical 2-D U-Net, story

3-D U-Net on an MRI-like volume¶

[B, C, X, Y, Z] as a first-class object: anatomy → translucent voxel blocks → a 3-D segmentation.

3-D U-Net, cinematic

3-D U-Net, technical 3-D U-Net, story 3-D U-Net, projection mode

python examples/medical_3d.py --mode volume|ortho|montage|projection

3-D transformer U-Nets on a whole head¶

MONAI UNETR and Swin UNETR, trained here on 64³ windows of a synthetic head with five structures. The transformer levels are stages; the output card is the whole head fused from sliding windows.

UNETR, cinematic

Swin UNETR, cinematic

UNETR, optional flat 2-D view

3-D U-Net on thick slices, drawn to scale

python examples/transformer_3d.py [--model swinunetr] [--flat] [--movie]
neural-flow render unest -i T1_mni.nii.gz --sliding-window --style cinematic      # MASI UNesT, real weights

See 3-D models.

nnU-Net: TotalSegmentator on a CT¶

Real TotalSegmentator weights (nnU-Net v2) on its public example CT, with nnU-Net's own preprocessing and sliding windows. Encoder levels bridge to the decoder levels of the same resolution.

TotalSegmentator organs, 1.5 mm, 27 windows

TotalSegmentator total, fast 3 mm model, 117 structures

neural-flow fetch totalseg totalseg-organs
neural-flow render totalseg-organs -i sample:ct --sliding-window --style cinematic
neural-flow render nnunet:path/to/results_folder -i case.nii.gz --sliding-window   # your nnU-Net