Home / AI Models / Finetuned-DINOv2-Chest-CT
Model released feature-extraction cc-by-nc-sa-4.0

Finetuned-DINOv2-Chest-CT

🤗 Hugging Face

Details

Architecture ViT-Large with Registers (1024-dim, patch 14)
Base Model DINOv2 ViT-Large with Registers
Relation finetune
License cc-by-nc-sa-4.0

DINOv2 ViT-Large with register tokens, finetuned on the CT-RATE dataset for chest CT feature extraction. Uses anatomically-aware cropping strategy that extracts 2D slices from 12mm physical slabs to preserve 3D spatial context.

Architecture

  • Embedding dimension: 1024
  • Patch size: 14
  • Register tokens: 4
  • Native resolution: 518x518
  • Input: 1-channel grayscale CT

Training

  • Hardware: 8x NVIDIA H200 GPUs
  • Strategy: FSDP bf16
  • Iterations: 20,000
  • Batch size: 48 per GPU
  • Data: CT-RATE train split (~120k samples)
  • Preprocessing: HU clipping [-997, 888], Z-score normalization (mean=-142.39, std=360.97)