DINOv3 with Test-Time Calibration for Automated Carotid Intima-Media Thickness Measurement on CUBS v1

1. Problem Statement

Carotid Intima-Media Thickness (CIMT) is a critical biomarker for assessing vascular aging, atherosclerosis, and cardiovascular risk. Traditionally, CIMT is measured manually from B-mode ultrasound by identifying the lumen-intima (LI) and media-adventitia (MA) interfaces. However, manual tracing is labor-intensive and suffers from inter- and intra-reader variability.

While numerous automated methods (edge detectors, active contours, and deep learning models) have been proposed, existing solutions often face two main challenges:

1. Lack of Transferability: Many models are highly specialized or dataset-specific, lacking the robustness of modern foundation models.
2. Measurement vs. Overlap: Most deep learning approaches optimize for segmentation overlap (e.g., Dice score) rather than the physical accuracy of the final thickness measurement. A model can have high overlap but still produce clinically significant errors in thickness due to boundary offsets or incorrect unit conversion.

The authors aim to bridge this gap by leveraging Vision Foundation Models (specifically DINOv3) and introducing a calibration-aware pipeline that prioritizes physical measurement accuracy over simple pixel overlap.

2. Methodology

The proposed framework consists of four main stages:

A. Dataset and Preprocessing

• Dataset: The study utilizes the CUBS v1 (Carotid Ultrasound Boundary Study) dataset, containing 2,176 longitudinal B-mode images from 1,088 participants.
• Ground Truth: Uses "Manual-A1" expert tracings as the gold standard.
• Splitting: To prevent data leakage, the dataset is split at the patient level (not image level) using three random seeds (42, 123, 999) for training, validation, and testing.
• Input Processing: Images are converted to grayscale, resized to $512 \times 512$, and normalized using ImageNet statistics.

B. DINOv3 Segmentation Model

• Architecture: The backbone is a DINOv3 encoder (vit_base_patch16_dinov3 from the timm ecosystem), followed by a lightweight convolutional segmentation head.
• Training Objective: The model is trained to minimize a combination of Cross-Entropy and Dice Loss.
• Goal: To generate a binary mask of the intima-media band.

C. Resize-Aware Unit Correction

A critical technical contribution is the handling of physical units. Since the network operates on resized images ($512 \times 512$) but the ground truth calibration factors (mm/pixel) are defined for the original resolution:

1. The system extracts the original calibration factor ($CF_{orig}$) and original image height ($H_{orig}$).
2. It calculates the corrected pixel size for the resized grid:
   $$\text{mm/pixel}_{512} = \text{mm/pixel}_{orig} \times \frac{H_{orig}}{512}$$
3. The thickness in pixels ($t_{px}$) is converted to micrometers ($\mu m$) using this corrected factor:
   $$t_{\mu m} = t_{px} \times \text{mm/pixel}_{512} \times 1000$$

D. Test-Time Calibration

Instead of using a fixed threshold (e.g., 0.5) to convert probability maps to binary masks, the authors introduce a test-time calibration step:

• The threshold is swept on a validation subset.
• The optimal threshold is selected specifically to minimize the CIMT absolute error, rather than maximizing segmentation overlap.
• This aligns the decision rule directly with the downstream clinical measurement objective.

3. Key Contributions

1. Foundation Model Adaptation: Successfully adapted a DINOv3-based architecture for the specific task of carotid intima-media complex segmentation, demonstrating its viability in medical imaging.
2. Physical Unit Fidelity: Developed a rigorous pipeline that propagates original image calibration factors through the resizing process, ensuring measurements are reported in physical units ($\mu m$) rather than arbitrary pixel counts.
3. Measurement-Oriented Calibration: Introduced a test-time threshold calibration strategy that optimizes for the final clinical metric (CIMT error) rather than segmentation overlap (Dice/IoU), significantly reducing systematic bias.

4. Results

The model was evaluated across three patient-level test splits:

• Segmentation Performance:

  • Mean Test Dice: 0.7739 ± 0.0037
  • Mean Test IoU: 0.6384 ± 0.0044
  • The model demonstrated stable performance despite the thin structure and low contrast of carotid boundaries.

• CIMT Measurement Accuracy:

  • Mean Absolute Error (MAE): 181.16 ± 11.57 µm (approx. 0.18 mm).
  • Mean Bias: 1.74 ± 37.07 µm (indicating near-zero systematic bias on average).
  • Pearson Correlation: 0.480 ± 0.259 (variable across splits).

• Impact of Test-Time Calibration:

  • On a held-out validation subset ($n=28$), the default threshold (0.5) yielded an MAE of 141.0 µm with a high bias of 101.3 µm.
  • After calibration (optimal threshold $\approx 0.80$), the MAE dropped to 101.1 µm, and the bias was reduced to -14.9 µm.
  • This demonstrates that post-processing calibration can substantially improve clinical accuracy even without retraining the model.

• Comparison: The results place the DINOv3-based approach within the clinically relevant ~0.1 mm measurement regime, comparable to the best classical computerized methods (TUMDE and CNR IT) reported in the original CUBS benchmark.

5. Significance and Conclusion

This work establishes that Vision Foundation Models can be effectively utilized for interpretable, physically calibrated medical measurements. The study highlights a crucial methodological insight: segmentation overlap (Dice) and thickness accuracy are not identical. A model can achieve high overlap while still producing clinically erroneous thickness measurements due to boundary shifts.

By explicitly correcting for image resizing and optimizing the inference threshold for the specific measurement endpoint, the authors achieved a robust CIMT estimation pipeline. This approach offers a modern, reproducible alternative to classical methods, paving the way for foundation models to be integrated into clinical workflows where physical accuracy is paramount.

Limitations noted by the authors:

• Generalization to other carotid ultrasound cohorts (external validation) is not yet established.
• Correlation scores varied across seeds, suggesting sensitivity to data splits.
• Current calibration is limited to threshold tuning rather than full self-supervised test-time adaptation of the model weights.