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

MedNeXt-v2: Scaling 3D ConvNeXts for Large-Scale Supervised Representation Learning in Medical Image Segmentation Saikat Roy∗,1,2, Yannick Kirchhoff∗,1,2,3, Constantin Ulrich1,4,5, Maximillian Rokuss1,2, Tassilo Wald1,2,6, Fabian Isensee1,6, Klaus Maier-Hein1,2,7 1German Cancer Research Center (DKFZ) Heidelberg, Division of Medical Image Computing, Germany 2Faculty of Mathematics and Computer Science, Heidelberg University, Germany 3HIDSS4Health - Helmholtz Information and Data Science School for Health, Karlsruhe/Heidelberg, Germany 4Medical Faculty Heidelberg, Heidelberg University, Germany 5National Center for Tumor Diseases (NCT), Heidelberg, Germany 6Helmholtz Imaging, German Cancer Research Center, Germany 7Pattern Analysis and Learning Group, Department of Radiation Oncology, Heidelberg University Hospital, Germany saikat.roy@dkfz-heidelberg.de; yannick.kirchhoff@dkfz-heidelberg.de Abstract Large-scale supervised pretraining is rapidly reshaping 3D medical image segmentation. However, existing efforts fo- cus 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 back- bones in large-scale pretraining pipelines are often subopti- mal. Subsequently, we use comprehensive backbone bench- marking prior to scaling and demonstrate that stronger from scratch performance reliably predicts stronger down- stream performance after pretraining. Guided by these find- ings, we incorporate a 3D Global Response Normalization module and use depth, width, and context scaling to im- prove our architecture for effective representation learning. We pretrain MedNeXt-v2 on 18k CT volumes and demon- strate state-of-the-art performance when fine-tuning across six challenging CT and MR benchmarks (144 structures), showing consistent gains over seven publicly released pre- trained models. Beyond improvements, our benchmarking of these models also reveals that stronger backbones yield better results on similar data, representation scaling dispro- portionately benefits pathological segmentation, and that modality-specific pretraining offers negligible benefit once full finetuning is applied. In conclusion, our results es- tablish 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. 1. Introduction Automated segmentation of medical images is one of the most common tasks in biomedical image analysis [12, 28, 49, 52]. Despite rapid development in deep learning based approaches over the last decade [1, 42, 50], UNet-based [59] deep convolutional networks (ConvNets) have remained central to high-performing methodologies for 3D medical image segmentation [31, 32]. Although alternative ap- proaches such as Transformers have been popular in recent years [36], their limited inductive bias has proved a hin- drance for training from scratch on the currently available medical segmentation datasets, typically containing sparse annotations [42]. This has led to ConvNeXt-based [46] approaches leveraging the scalability of the Transformer *Contributed equally. Each author may denote themselves as posi- tional first author in their CVs. 1 arXiv:2512.17774v1 [eess.IV] 19 Dec 2025 SegVol Vista3D TotalSeg CADS MedNeXt-v1 MedNeXt-v2 (ours) 70 72 74 76 78 80 82 84 Mean DSC over Datasets +11.33 +2.69 +2.50 +1.15 +1.23 Figure 1. MedNeXt-v2 sets a new state-of-the-art in 3D medical image segmentation. By leveraging micro-architectural improve- ments and large-scale pretraining, it outperforms powerful existing networks across multiple 3D medical segmentation tasks. while retaining the strong inductive bias of ConvNets to of- fer effective solutions for 3D medical image segmentation [10, 38, 43, 56, 61]. However, following significant advances in computer vi- sion [13, 55, 66] over the last decade, the field of med- ical image segmentation has also been gradually mov- ing towards incorporating large-scale supervised pretrain- ing of deep networks [9, 26, 68, 72]. In recent years, the availability of large monolithic datasets [19, 73] or aggregated collections of previously available small-scale datasets [3, 39, 76] has led to initial attempts at pretrain- ing large-scale deep learning models for the segmentation of 3D medical images. Notably, while approaches in 2D com- puter vision have moved towards self-supervised learning (SSL) owing to the abundance of unlabeled data [23, 34], the domain of 3D medical image segmentation continues to leverage supervised pretraining. Despite s

This content is AI-processed based on open access ArXiv data.