Home / AI Models / Guided-Chest-CT-LeJEPA
Model released feature-extraction cc-by-nc-sa-4.0

Guided-Chest-CT-LeJEPA

🤗 Hugging Face

Details

Architecture ViT-Large (1024-dim, patch 14)
Relation trained-from-scratch
License cc-by-nc-sa-4.0

Extension of Chest-CT-LeJEPA with semi-3D cropping and auxiliary supervised loss. The key innovation is anatomical guidance: 80% of local crops are centered on anatomical structures from 118 TotalSegmentator organ classes, forcing the model to learn clinically relevant features.

Architecture

  • Base: ViT-Large (vit_large_patch14_dinov2 via timm)
  • Embedding dimension: 1024
  • Patch size: 14
  • Input: 224x224 global crops, 140x140 local crops
  • Auxiliary loss: Classification head over 118 anatomical structure labels

Training

  • Hardware: 8x NVIDIA H200 GPUs
  • Strategy: DDP
  • Iterations: 50,000 (2 phases of 25,000 steps)
  • Global batch size: 512
  • Data: CT-RATE + ReXGroundingCT
  • Anatomical guidance: 80% of local crops centered on structures

This model serves as the visual encoder for the Ker-VLJEPA-3B radiology report generation system.