Anatomy crosses slice boundaries
Cardiac structures extend across the scan. Segmenting each slice independently can lose the geometry that connects them.
Volumetric vision / segmentation
The notebooks explore hierarchical transformer representations for three-dimensional anatomy, with voxel-level masks as the output.
Interactive architecture
The encoder processes overlapping 3D patches at four scales. Features from all four scales feed the segmentation decoder.
Case study
Cardiac structures extend across the scan. Segmenting each slice independently can lose the geometry that connects them.
A MixTransformer-style hierarchy takes overlapping 3D patch embeddings and produces features at multiple resolutions. The decoder uses those features to predict voxel labels.
The public repository follows the work through notebooks: loading data, inspecting segmentations, building the model, and running training experiments. It also includes checkpoints.
Architecture detail
Keep the scan slices and target masks together as one three-dimensional sample.
Use overlapping patches to preserve boundaries that would otherwise fall at rigid token edges.
Each stage trades some spatial resolution for a broader semantic representation.
Combine features from each stage to recover a dense voxel-level segmentation.
Evidence boundary
This is an implementation experiment, not a clinical system. The notebooks document the architectural exploration and training workflow; they do not establish prospective clinical performance, deployment readiness, or medical-device validation.