DINO-3DRA

Abstract

Accurate aneurysm segmentation in 3D rotational angiography (3DRA) is hindered by extreme class imbalance, morphological similarity to vessels, and absent large-scale 3D pretraining. 2D vision foundation models encode dense structural priors from ~1.7 billion images, yet naïve slice-wise transfer fragments anatomical continuity and destabilises optimisation.

We propose DINO-3DRA, a dual-path framework achieving effective cross-dimensional semantic transfer by injecting frozen DINOv3 features into a 3D U-Net via Room-Lite spatial mixing and calibrated residual fusion. On multi-centre 3DRA data, DINO-3DRA achieves state-of-the-art aneurysm segmentation (Dice: 0.758; HD95: 2.75 mm; +13% over nnU-Net) with only 5.72M trainable parameters. Ablation studies confirm that gains arise from structured cross-dimensional transfer rather than loss design alone. Cross-dataset evaluation eliminates all catastrophic failure cases, demonstrating robust generalisation across heterogeneous imaging protocols.

The Challenge

Ambiguity. 3DRA produces near-binary contrast — bright vessel lumen on dark background. On any single 2D slice, both a vessel cross-section and an aneurysm appear identical: a white circle on black. Distinguishing them requires observing how the cross-sectional shape changes across consecutive depth slices — a vessel maintains a consistent tube, while an aneurysm bulges and contracts.

The 2D–3D gap. The best 2D models (DINOv3) are trained on 1.7B images with rich structural priors. No equivalent large-scale 3D pretraining exists. But naïvely injecting 2D features into 3D models fails — it makes performance worse than using no 2D features at all.

Method

DINO-3DRA pairs a frozen 2D VFM with a trainable 3D U-Net. The frozen branch provides what kind of anatomy is in each slice. The trainable branch provides where exactly structures are. Two lightweight modules bridge the gap.

1 — Slice-Level Semantic Extraction. All 64 axial slices processed independently by frozen DINOv3-Small. 196 patch tokens globally averaged to a single 384-dim semantic embedding per slice. Projected via 1×1 convolutions and stacked into 3D volumes at three U-Net skip scales.

(B×64, 196, 384) → mean-pool → (B×64, 384) → (B, {24/48/96}, {64³/32³/16³})

2 — Room-Lite Depth Coherence. Learnable depth-positional bias B(z) + depthwise-separable 3D convolutions restore inter-slice continuity. Without this, Dice collapses −49%. 21K params total.

Y = X + GELU(GN(Conv₁ₓ₁(Conv₃ₓ₃(X + B(z)))))

3 — Calibrated Residual Fusion. GN-normalised DINO features injected as a learnable residual correction into U-Net skips. The residual bypass guarantees DINO can only help, never hurt.

F_fused = F_U + GELU(GN(W[F_U; GN(F_D)]))

Results

Ablation Study

Four of five DINO variants fall below the plain U-Net baseline. Only the full pipeline — Room-Lite + Calibrated Fusion — surpasses it.

Cross-Dataset Generalisation

Evaluated on two external datasets without fine-tuning. DINO-3DRA eliminates all catastrophic failures (Dice < 0.5).

BibTeX

@inproceedings{dino3dra2026,
  title     = {DINO-3DRA: Leveraging 2D Foundation Model
               Semantics for 3D Cerebral Aneurysm Segmentation},
  author    = {Anonymized Authors},
  booktitle = {International Conference on Medical Image Computing
               and Computer-Assisted Intervention (MICCAI)},
  year      = {2026}
}