MedNeXt-v2: Scaling 3D ConvNeXts for Large-Scale Supervised Representation Learning in Medical Image Segmentation (2512.17774v1)

Abstract: Large-scale supervised pretraining is rapidly reshaping 3D medical image segmentation. However, existing efforts focus primarily on increasing dataset size and overlook the question of whether the backbone network is an effective representation learner at scale. In this work, we address this gap by revisiting ConvNeXt-based architectures for volumetric segmentation and introducing MedNeXt-v2, a compound-scaled 3D ConvNeXt that leverages improved micro-architecture and data scaling to deliver state-of-the-art performance. First, we show that routinely used backbones in large-scale pretraining pipelines are often suboptimal. Subsequently, we use comprehensive backbone benchmarking prior to scaling and demonstrate that stronger from scratch performance reliably predicts stronger downstream performance after pretraining. Guided by these findings, we incorporate a 3D Global Response Normalization module and use depth, width, and context scaling to improve our architecture for effective representation learning. We pretrain MedNeXt-v2 on 18k CT volumes and demonstrate state-of-the-art performance when fine-tuning across six challenging CT and MR benchmarks (144 structures), showing consistent gains over seven publicly released pretrained models. Beyond improvements, our benchmarking of these models also reveals that stronger backbones yield better results on similar data, representation scaling disproportionately benefits pathological segmentation, and that modality-specific pretraining offers negligible benefit once full finetuning is applied. In conclusion, our results establish MedNeXt-v2 as a strong backbone for large-scale supervised representation learning in 3D Medical Image Segmentation. Our code and pretrained models are made available with the official nnUNet repository at: https://www.github.com/MIC-DKFZ/nnUNet

• The paper introduces MedNeXt-v2, a novel compound-scaled 3D ConvNeXt architecture that integrates GRN modules for stabilizing training in medical segmentation tasks.
• It employs depth, width, and context scaling, achieving up to +1.0 DSC improvement over state-of-the-art baselines across CT and MR datasets.
• The work emphasizes rigorous backbone selection and systematic benchmarking, enabling efficient representation learning and reliable transfer to clinical segmentation tasks.

MedNeXt-v2: Advancing Scalable Supervised Representation Learning for 3D Medical Image Segmentation

Introduction

The domain of 3D medical image segmentation has witnessed rapid shifts toward large-scale supervised representation learning, paralleling developments in general computer vision. However, many efforts remain constrained by legacy backbones and limited architectural benchmarking, with a prevailing focus on expanding dataset size. The paper "MedNeXt-v2: Scaling 3D ConvNeXts for Large-Scale Supervised Representation Learning in Medical Image Segmentation" (2512.17774) addresses systemic limitations by introducing MedNeXt-v2—a compound-scaled 3D ConvNeXt architecture designed for state-of-the-art representation learning. This work performs rigorous backbone selection prior to pretraining, integrates micro-architectural advances such as 3D Global Response Normalization (GRN), and systematically benchmarks MedNeXt-v2 against prominent public pretrained networks across diverse segmentation challenges.

Figure 1: MedNeXt-v2 sets a new state-of-the-art in 3D medical image segmentation, outperforming leading baselines across multiple 3D segmentation tasks.

Architectural Innovations and Scaling Strategies

MedNeXt-v2 is built upon the ConvNeXt foundation but compounds scaling across depth, width, and input context. The architecture incorporates 3D GRN modules to mitigate activation collapse prevalent in deep ConvNets, thereby fostering robust feature diversity throughout volumetric representation learning.

Figure 2: Micro-architecture of MedNeXt-v2, highlighting the integration of GRN and channel scaling strategies.

Noteworthy compounds of scaling in MedNeXt-v2 include:

• Depth scaling: Adoption of a deep 52-layer configuration, paralleling efficient scaling paradigms akin to EfficientNet.
• Width scaling: Doubling channel capacity at each network stage to enhance representational power.
• Context scaling: Fine-tuning with enlarged $192^3$ input patches amplifies spatial context, leading to notable downstream accuracy gains despite modest pretraining resource demands.

GRN is shown to consistently stabilize training dynamics and prevent dead or saturated channel activations. Channel activation visualization further evidences the role of GRN in sustaining informative, non-redundant features across varied anatomical and pathological targets.

Figure 3: Channel activation visualization—3D GRN in MedNeXt-v2 reduces redundant activations, preventing feature collapse.

Empirical Evaluation and Benchmarking

MedNeXt-v2 undergoes benchmarking on six challenging public datasets, totaling 144 structures spanning CT and MR modalities. Pretraining is conducted on 18,000 CT volumes covering 44 target structures, paired with rigorous fine-tuning regimens. The design enforces backbone validation against state-of-the-art baselines prior to scaling, ensuring optimal representation capabilities.

The incorporation of GRN distinctly improves performance both when training from scratch and after large-scale supervised pretraining. MedNeXt-v2 outperforms MedNeXt-v1 and alternative pretrained baselines—such as nnUNet, ResEncL, Vista3D, TotalSegmentator, MRSegmentator, and STU-Net—by up to +1.0 DSC on difficult datasets, including pediatric CT and pancreatic tumor segmentation.

Figure 4: MedNeXt-v2 outperforms MedNeXt-v1 from scratch and during fine-tuning, attributed to GRN stabilization.

Further, context scaling during fine-tuning yields substantial benefit, with larger patches enabling superior segmentation at anatomical boundaries and improving performance metrics relative to both smaller patch variants and competing public models.

Figure 5: Increased context during fine-tuning with $192^3$ patches enables more accurate segmentation near image boundaries.

Qualitative analysis demonstrates MedNeXt-v2's capability in both CT and MR domains, with notable improvements in boundary delineation and fine anatomical detail.

Figure 6: Qualitative visualizations of MedNeXt-v2 and context-scaled variant on CT and MR datasets, compared against public baselines.

Insights from Systematic Benchmarking

Comprehensive benchmarking leads to several impactful findings:

• Superior backbone yields superior pretraining: Downstream fine-tuning performance reliably tracks from-scratch validation; architectures with higher initial representational power benefit most from large-scale pretraining.
• Pathological segmentation gains: Tumor and lesion classes exhibit disproportionate improvement from pretraining, outperforming healthy tissue segmentation benchmarks.
• Modality-agnostic pretraining: Pretraining on CT or MR modalities translates well across both domains, suggesting structurally similar modalities benefit from shared, generalized representations.
• Scaling limits: Network capacity scaling delivers mixed returns; larger architectures require careful dataset scaling and may not outperform more context-aware or architecturally optimized models, particularly in saturated benchmarks.

  Figure 7: Tumor classes benefit massively in pretraining—pretrained networks yield significant gains for pathological segmentation vs organ classes.

MedNeXt-v2 Macro Architecture

MedNeXt-v2 features a symmetric 5-hierarchy UNet-style architecture with compound scaling and residual ConvNeXt blocks at each stage. GRN is integrated throughout all residual paths and up/downsampling layers for effective learning dynamics.

Figure 8: MedNeXt-v2 macro-architecture parallels MedNeXt-v1, but augments representation via GRN and scalable design.

Implications and Future Directions

MedNeXt-v2 conclusively demonstrates that strategic backbone selection, combined with sensible micro-architectural innovation, is pivotal for the advancement of large-scale 3D medical image segmentation models. The findings prompt a methodological shift from dataset size-centric practices toward representation-aware architecture benchmarking and scaling. Practical implications include more reliable transfer learning for downstream clinical segmentation tasks and rapid fine-tuning pipelines with minimized resource footprint.

Theoretically, the work substantiates the role of architectural bias—such as ConvNeXt-style inductive properties—over mere parameter count or data scale, particularly in data-constrained and annotation-sparse regimes characteristic of medical imaging. Future research may explore adaptive scaling strategies in response to domain-specific saturation, further cross-modality generalization, or integration with semi/self-supervised pipelines as annotation burdens persist.

Conclusion

MedNeXt-v2 sets a new standard in 3D supervised medical image segmentation, combining rigorous backbone validation, micro-architectural refinement via GRN, and efficient scaling strategies. Its systematic benchmarking against state-of-the-art pretrained models reaffirms the importance of representation-centric design. The work lays a robust foundation for optimized segmentation pipelines and motivates continued research in scalable, adaptive architectures for medical AI.