Volumetric vision / segmentation

Cardiac MRI segmentation.

The notebooks explore hierarchical transformer representations for three-dimensional anatomy, with voxel-level masks as the output.

Interactive architecture

From 3D patches back to voxels.

The encoder processes overlapping 3D patches at four scales. Features from all four scales feed the segmentation decoder.

Case study

Keeping the whole volume in view

01 / Problem

Anatomy crosses slice boundaries

Cardiac structures extend across the scan. Segmenting each slice independently can lose the geometry that connects them.

02 / System

Encode the volume at several scales

A MixTransformer-style hierarchy takes overlapping 3D patch embeddings and produces features at multiple resolutions. The decoder uses those features to predict voxel labels.

03 / Artifact

Notebooks and checkpoints

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

From volume to mask

  1. Load the volume

    Keep the scan slices and target masks together as one three-dimensional sample.

  2. Tokenize in 3D

    Use overlapping patches to preserve boundaries that would otherwise fall at rigid token edges.

  3. Expand context

    Each stage trades some spatial resolution for a broader semantic representation.

  4. Decode scales

    Combine features from each stage to recover a dense voxel-level segmentation.

Evidence boundary

An architectural experiment.

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.