DALE-CT
Overview
Training code and model release for DALE-CT (Depth-Aware Latent-Euclidean Computed Tomography) — a family of 2D slice-based Vision Transformers trained entirely from scratch on chest CT with the heuristics-free LeJEPA objective and depth-aware slab sampling: self-supervised views are drawn from across a physical z-axis slab rather than a single slice, so the frozen representations form an anatomical world model — they linearly decode volumetric slice position (R^2 ≈ 0.97), recover slice ordering without labels, and localize organs and findings, despite no 3D or positional supervision.
Paper: DALE-CT: Depth-Aware 2D Slice Encoders Learn an Anatomical World Model of Chest CT · Benchmark: chest-ct-foundation-model-benchmark — the single-protocol evaluation harness, patient-disjoint splits, and full result tables for every model in the paper
From the project’s README on GitHub.
At a glance
| Repository | Kentucky-Open-Science/DALE-CT |
|---|---|
| Research area | Medical imaging & digital pathology |
| Primary language | Python |
| Languages | Python 99.3%, Dockerfile 0.7% |
| License | Apache-2.0 |
| Stars / forks | 0 / 0 |
| Open issues and pull requests | 0 |
| Created | 2026-05-08 |
| Last push | 2026-08-26 |
| Default branch | main |
| Topics | medical-imaging |
What the README covers
Top contributors
- @EvanDamron (18 commits)
Related projects
-
chest-ct-foundation-model-benchmark
Single-protocol benchmark for chest-CT foundation models: one frozen-probe protocol, shared patient-disjoint splits, bootstrap CIs, every encoder probed by the same harness
-
CLASSify-Connect
CLASSify is a self-service machine learning platform developed by Aaron Mullen for the Center for Applied Artificial Intelligence located at the University of Kentucky. The …
-
ECG_Predictor
Archived
Predict them beats
-
WebSAM-React
Archived
No description on GitHub yet.
Get the code
git clone https://github.com/Kentucky-Open-Science/DALE-CT.git