Skip to content

Changelog

0.3.0 (2026-09-30)

nnU-Net.

  • nnU-Net U-Nets level by level. Networks built by dynamic_network_architectures (PlainConvUNet, ResidualEncoderUNet) are recognised by structure and drawn as a complete U: stem, encoder levels, bottleneck, one stage per decoder level (upsampling and skip concatenation folded into its edges) and the full-resolution segmentation layer as the only head. Previously the decoder was mostly missing.
  • Deep supervision off at inference. Networks with deep_supervision=True are run with it switched off (and restored), so only the full-resolution prediction is drawn; aux_outputs=True keeps the rest.
  • Trained nnU-Net models. nnunet:RESULTS_DIR[:FOLD] (CLI and zoo.load_model) rebuilds the network from plans.json (old and current formats), loads the fold's checkpoint and applies nnU-Net's preprocessing (RAS, cropping, CT / z-score normalisation, resampling), patch size, spacing and 50 % window overlap. nnU-Net itself is not required.
  • TotalSegmentator. totalseg, totalseg-6mm and totalseg-organs download its public CT models; sample:ct its example CT. Tested against TotalSegmentator's own reference segmentation (Dice 0.95–0.98).
  • New example examples/nnunet_totalseg.py (figures and a sliding-window movie), docs page nnU-Net and TotalSegmentator, tests tests/test_nnunet.py.
  • volume_axes="zyx" for nnU-Net / SimpleITK order.
  • Many-label 3-D segmentations: only structures that enclose others (scalp, skull, white matter) are drawn as glass, so solid organs stay opaque; the whole-volume output card grows with the depth of the U.
  • Gradient explanations are skipped with a note, instead of running out of memory, when their backward pass would not fit (explain_max_mb).

0.2.0 (2026-09-29)

3-D transformer networks and whole-volume inference.

  • Transformer U-Nets as they are. UNETR, Swin UNETR, UNesT and similar networks are recognised from the runtime dataflow: the transformer's tapped blocks / levels become stages (patch embedding → transformer encoder → bottleneck → decoder → head), projection branches become skip bridges, and the figure is drawn as a U. Previously the backbone was drawn as one box (UNETR) or left out (Swin UNETR).
  • Voxel spacing. voxel_spacing=(sx, sy, sz); the command line reads it from the NIfTI header. Volumes, slices and segmentations are drawn to physical scale.
  • Whole-volume output. sliding_window=True, roi_size=…: the stages show the one window the network processed, the output card shows the prediction fused from every window (Gaussian weighting; memory-bounded for many-class models), with the traced window outlined.
  • 3-D inference movie. sliding_window_movie(...) / neural-flow movie MODEL --inference scan.nii.gz.
  • Optional flat view. flat_3d=True draws 3-D stages as squashed 2-D projections (off by default).
  • Many-label segmentations (e.g. 133 brain structures) get distinct colours, and large structures are drawn as glass so small, deep ones stay visible.
  • neural-flow fetch unest falls back to the MONAI model-zoo sources on GitHub plus the NVIDIA weight URL when Hugging Face / NGC are unreachable; bundle structure names are used as class names.
  • New example examples/transformer_3d.py (UNETR and Swin UNETR trained on a synthetic whole-head phantom with five structures), medical_3d.py --thick, rewritten monai_bundle.py.
  • New docs: 3-D models, About; references for every demo model.
  • Fix: animate_inputs(..., explain=...) raised a duplicate-keyword error.

0.1.0 (2026-09-29)

First public release.

  • visualize_model, trace_model, draw: representation-flow figures for arbitrary PyTorch models (CNN, U-Net, ViT/Swin, 3-D medical, multi-input / multi-head).
  • Automatic stage selection with explicit overrides (names, regex, type, predicate).
  • Runtime dataflow topology (skips, merges, branches) with torch.fx and execution-order fallbacks.
  • Memory-aware, on-device activation reduction; deterministic channel ranking and PCA.
  • Styles: technical, story, cinematic (black background, feature-map stacks, PCA front pages, beams, receptive-field circles, contribution lines, Grad-CAM); themes light, dark, black.
  • First-class 3-D volumes: cut-away anatomy, voxel blocks, activation-driven orthogonal slices, projections, 3-D segmentation renders.
  • Attention capture (MHA / SDPA) and attention rollout.
  • Interactive HTML explorer; "light-up" animation; movies over changing inputs (animate_inputs) with fixed channels, colours and scales; input-sequence generators.
  • neural-flow command-line tool: demo, render, movie, inspect, fetch, models.
  • Examples with real pretrained models (ResNet-50, ViT-B/16, TorchXRayVision DenseNet-121), demo models trained on synthetic data, a MONAI-bundle runner (MASI UNesT), and an example slide deck.