How aiconslab/3dino works
3DINO is a 3D medical-imaging self-supervised pretraining framework, with a pretrained 3D ViT intended for reuse in downstream tasks. The main workflow prepares unlabeled volumes, trains a student–teacher model, then uses pretrained features for classification or fine-tunes segmentation heads. The README also documents high-resolution adaptation and PCA visualization. Source excerpts and the index support the training and evaluation entry points; some internal data flow and unsampled evaluation details are shown only where documented.
Main entry points: segmentation3d.py Segmentation fine-tuning · ssl_meta_arch.py SSL training architecture · loaders.py Dataset loaders · linear3d.py Linear classification
Components 25
Medical image data
- Pretraining augmentation
dinov2/data/augmentations.py - Dataset loaders
dinov2/data/loaders.py - Data samplers
dinov2/data/samplers.py - Image transforms
dinov2/data/transforms.py
Self-supervised pretraining
- Training entry point
dinov2/train/train3d.py - SSL training architecture
dinov2/train/ssl_meta_arch.py - DINO class loss
dinov2/loss/dino_clstoken_loss.py - iBOT patch loss
dinov2/loss/ibot_patch_loss.py - KoLeo loss
dinov2/loss/koleo_loss.py
3DINO model
- 3D Vision Transformer
dinov2/models/vision_transformer.py - Transformer blocks
dinov2/layers/block.py - Attention layer
dinov2/layers/attention.py - MLP layer
dinov2/layers/mlp.py - 3D patch embedding
dinov2/layers/patch_embed3d.py
Downstream evaluation
- Evaluation setup
dinov2/eval/setup.py - Linear classification
dinov2/eval/linear3d.py - Segmentation fine-tuning
dinov2/eval/segmentation3d.py - Segmentation heads
dinov2/eval/segmentation_3d/segmentation_heads.py - ViT adapter
dinov2/eval/segmentation_3d/vit_adapter.py - Segmentation augmentation
dinov2/eval/segmentation_3d/augmentations.py - Segmentation metrics
dinov2/eval/segmentation_3d/metrics.py - Feature visualization
dinov2/eval/vis_pca.py
Other components
- Researcher
- Medical volumes
- Pretrained weights
Connections 26
- Researcher → Training entry point: starts training
- Medical volumes → Dataset loaders: provides dataset
- Dataset loaders → Data samplers: uses sampler
dinov2/data/loaders.py - Training entry point → SSL training architecture: trains model
dinov2/train/train3d.py - Training entry point → Dataset loaders: sets up data
dinov2/train/train3d.py - SSL training architecture → 3D Vision Transformer: builds backbone
dinov2/train/ssl_meta_arch.py - SSL training architecture → DINO class loss: computes class loss
dinov2/train/ssl_meta_arch.py - SSL training architecture → iBOT patch loss: computes patch loss
dinov2/train/ssl_meta_arch.py - SSL training architecture → KoLeo loss: computes KoLeo loss
dinov2/train/ssl_meta_arch.py - 3D Vision Transformer → Transformer blocks: constructs blocks
dinov2/models/vision_transformer.py - 3D Vision Transformer → 3D patch embedding: embeds patches
dinov2/models/vision_transformer.py - Transformer blocks → Attention layer: uses attention
dinov2/layers/block.py - Transformer blocks → MLP layer: uses MLP
dinov2/layers/block.py - Researcher → Linear classification: runs classification
- Researcher → Segmentation fine-tuning: runs segmentation
- Pretrained weights → Linear classification: loads weights
- Pretrained weights → Segmentation fine-tuning: loads weights
- Linear classification → Evaluation setup: configures evaluation
dinov2/eval/linear3d.py - Segmentation fine-tuning → Evaluation setup: configures evaluation
dinov2/eval/segmentation3d.py - Linear classification → Dataset loaders: builds dataset
dinov2/eval/linear3d.py - Segmentation fine-tuning → Dataset loaders: builds datasets
dinov2/eval/segmentation3d.py - Segmentation fine-tuning → Segmentation heads: selects head
dinov2/eval/segmentation3d.py - Segmentation heads → ViT adapter: uses adapter
dinov2/eval/segmentation_3d/segmentation_heads.py - Segmentation fine-tuning → Segmentation augmentation: uses transforms
dinov2/eval/segmentation3d.py - Segmentation fine-tuning → Segmentation metrics: evaluates metrics
dinov2/eval/segmentation3d.py - Researcher → Feature visualization: visualizes features