Do not treat slices as islands
Cardiac structures extend across the scan, and independently segmented slices can lose the geometry that makes the anatomy coherent.
Volumetric vision / segmentation
Cardiac MRI segmentation.
A notebook-led experiment adapting hierarchical transformer representations to three-dimensional anatomy and voxel-level masks.
Interactive architecture
A hierarchy that returns to voxel space.
Overlapping 3D patches move through four scales, then contribute together to the segmentation decoder.
Case study
Carrying context through a volume
Build a multi-scale volume
Overlapping 3D patch embeddings feed a MixTransformer-style hierarchy whose multiple resolutions are decoded into voxel predictions.
Keep experimentation visible
The public repository is notebook-led, covering data loading, segmentation inspection, model construction, training experiments, and checkpoints.
Architecture detail
From volume to mask
Load the volume
Organize scan slices and target masks as a coherent three-dimensional sample.
Tokenize in 3D
Use overlapping patches so boundaries do not vanish at rigid token edges.
Expand context
Move through hierarchical stages that trade spatial resolution for broader semantics.
Decode scales
Fuse features from each stage and recover a dense voxel-level segmentation.
Evidence boundary
An implementation experiment, not a clinical system.
The notebooks document an architectural exploration and training workflow. They do not establish prospective clinical performance, deployment readiness, or medical-device validation.