Resolution scaling governs DINOv3 transfer performance in chest radiograph classification

Jakob Nikolas Kather, Md, MSc, Soroosh Tayebi Arasteh, Mina Shaigan, Christiane Kuhl, Jakob Nikolas Kather, Sven Nebelung, Daniel TruhnarXiv 2025 · DOI · arXiv

Introduction

Self-supervised learning (SSL) has transformed computer vision, yet medical imaging—particularly chest X-rays (CXR)—presents unique challenges. Detecting fine-grained findings like tiny nodules or subtle textural changes requires high spatial resolution. While standard models are often pretrained at 224×224 pixels, Meta’s DINOv3 introduces architectural features specifically designed for high-resolution adaptation. This study benchmarks DINOv3 against its predecessors to determine if these improvements translate to clinical utility.

Experimental Design and Datasets

The researchers conducted a large-scale evaluation using over 814,000 images from seven diverse datasets, including Pedi-CXR, VinDr-CXR, and MIMIC-CXR. The study systematically compared:

  • Initializations: Supervised ImageNet-21K, DINOv2, and DINOv3.
  • Backbones: ViT-B/16 (Transformer) vs. ConvNeXt-B (Convolutional).
  • Resolution: Scaling from 224×224 up to 1024×1024 pixels.
  • Training Strategy: Full finetuning vs. frozen features from a 7B-parameter teacher model.
Overview of the benchmark datasets, resolutions, backbones, and pretraining strategies evaluated in the study.

Key Insights

Resolution as the Primary Performance Driver

A critical finding is that DINOv3’s advantages are resolution-dependent. At the standard 224×224 resolution, DINOv3 performs similarly to DINOv2. Its superior performance only emerges at 512×512 pixels, validating its high-resolution adaptation design. While scaling to 1024×1024 provided some gains, 512px was identified as the "sweet spot" for balancing accuracy and computational efficiency.

Backbone and Adaptation Strategy

Despite the popularity of Vision Transformers, the ConvNeXt-B backbone consistently outperformed ViT-B/16. The inductive biases of convolutional neural networks (CNNs) appear better suited for the structured spatial data in radiographs. Furthermore, domain adaptation proved more vital than model scale: a fully finetuned 89M-parameter model outperformed a frozen 7B-parameter model, emphasizing the need for task-specific tuning in radiology.

Results and Clinical Implications

The strongest configuration was a DINOv3-initialized ConvNeXt-B at 512×512 resolution. This setup achieved AUROC gains of 0.8–1.5 percentage points over ImageNet and DINOv2 baselines. These improvements were most significant for "boundary-dependent" findings, such as pneumothorax and pulmonary nodules, where fine spatial detail is paramount.

Interestingly, the pediatric cohort (Pedi-CXR) did not show the same performance gaps between initializations, likely due to smaller sample sizes and distinct anatomical differences compared to adult populations.

Conclusion

Resolution scaling is a fundamental factor in the transfer performance of modern SSL models like DINOv3. By leveraging 512×512 resolution and convolutional backbones, DINOv3 provides a meaningful step forward in the automated classification of complex chest radiograph findings. Future research should investigate whether these resolution-driven gains extend to localization and segmentation tasks across other high-resolution medical modalities.

Summary updated

Abstract

Self-supervised learning (SSL) has improved visual representation learning, but its value in chest radiography remains uncertain. DINOv3 extends earlier SSL models through Gram-anchored self-distillation and explicit high-resolution adaptation. Whether these changes improve transfer learning for chest radiograph classification has not been established. We benchmarked DINOv3 against DINOv2 and supervised ImageNet initialization across seven chest radiograph datasets comprising 816,183 radiographs from pediatric and adult cohorts. ViT-B/16 and ConvNeXt-B were evaluated under full fine-tuning at 224 and 512 pixels, with targeted 1024 experiments on three cohorts. Additional analyses examined parameter-efficient adaptation, synthetic label corruption, external validation, frozen 7B features, and computational efficiency. The primary outcome was mean AUROC across labels. In adult cohorts, DINOv3 did not consistently outperform DINOv2 at 224 x 224 pixels, but became the strongest initialization at 512 x 512, especially with ConvNeXt-B. Gains were greatest for small focal and boundary-dependent abnormalities, whereas large-structure findings changed little. The pediatric cohort showed no significant benefit from DINOv3, higher resolution, or backbone choice. Scaling to 1024 x 1024 rarely improved performance and markedly increased computational cost. ConvNeXt-B remained superior to ViT-B/16 under both full and parameter-efficient adaptation. External validation preserved the 512 x 512 DINOv3 advantage, whereas synthetic label corruption showed that this benefit should not be interpreted simply as superior noise robustness. For adult chest radiograph classification, DINOv3 provides its most reliable benefit at 512 x 512 pixels, particularly with ConvNeXt-B. Fully adapted mid-sized models at 512 x 512 pixels provided the best performance-cost trade-off in our benchmark.
BibTeX
@misc{msc2025resolution,
  author = {Jakob Nikolas Kather, MD, MSc and Soroosh Tayebi Arasteh and Mina Shaigan and Christiane Kuhl and Jakob Nikolas Kather and Sven Nebelung and Daniel Truhn},
  title = {{Resolution scaling governs DINOv3 transfer performance in chest radiograph classification}},
  year = {2025},
  eprint = {2510.07191},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.07191},
}
Figure