---
title: 'MedNeXt-L-k5: Large 3D Segmentation Model'
url: https://www.emergentmind.com/topics/mednext-l-k5
type: topic
---

# MedNeXt-L-k5: Large 3D Segmentation Model

Searching arXiv for MedNeXt and MedNeXt-L-k5 to ground the article in relevant papers.
MedNeXt-L-k5 denotes the Large MedNeXt configuration instantiated with \(5\times5\times5\) depthwise kernels for 3D medical image segmentation. In the original MedNeXt paper, the exact compact string is not printed verbatim; instead, the model appears as MedNeXt-L evaluated at kernel \(5\), within a compound-scaling scheme over block count, expansion ratio, and kernel size [2303.09975]. The term later appears explicitly in some downstream work, most directly in automated perivascular-space segmentation [2508.20256]. Across the subsequent literature, however, the label is not used uniformly: some papers use exact MedNeXt-L-k5 terminology, while others describe only a “large-kernel MedNeXt” that enlarges receptive field without using the Large-capacity MedNeXt-L scale [2508.01831].

## 1. Definition and nomenclature

In the MedNeXt family, the scale labels are \(S\), \(B\), \(M\), and \(L\), and kernel size is treated as a separate axis. Under this convention, “L” denotes the largest predefined MedNeXt configuration, and “k5” denotes the use of \(5\times5\times5\) depthwise kernels rather than \(3\times3\times3\) kernels [2303.09975]. The original work explicitly studies \(k \in \{3,5\}\), so MedNeXt-L-k5 is the large-scale, kernel-5 member of that experimental grid [2303.09975].

A recurring source of confusion is that “large” does not always mean the same thing in later papers. In the breast MRI study "Large Kernel MedNeXt for Breast Tumor Segmentation and Self-Normalizing Network for pCR Classification in Magnetic Resonance Images" [2508.01831], “large-kernel MedNeXt” refers to a receptive-field enlargement from \(3\times3\times3\) to \(5\times5\times5\), while the actual instantiated backbone is explicitly the **medium** variant of MedNeXt, not MedNeXt-L. By contrast, "MedNet-PVS: A MedNeXt-Based Deep Learning Model for Automated Segmentation of Perivascular Spaces" [2508.20256] uses the exact label MedNeXt-L-k5 and explains it as the **large** MedNeXt configuration with a \(5\times5\times5\) kernel.

A second misconception concerns model class. MedNeXt-L-k5 is routinely described as “Transformer-inspired,” but the architecture is not a transformer proper. The PVS paper characterizes MedNeXt as a “Transformer-inspired, fully ConvNeXt 3D Encoder-Decoder Network” and also as a “large-kernel ConvNeXt segmentation architecture,” which places it in the family of fully convolutional 3D encoder-decoder models rather than self-attention-based transformers [2508.20256].

## 2. Architectural specification

MedNeXt is defined as a fully ConvNeXt-style 3D encoder-decoder segmentation network with a UNet-like macro-architecture comprising **4 encoder layers**, **1 bottleneck layer**, and **4 decoder layers**, with skip connections between corresponding resolutions and additive residuals throughout [2303.09975]. Deep supervision is applied at each decoder layer, with lower loss weights at lower resolutions [2303.09975].

The basic MedNeXt block is a three-convolution inverted bottleneck. Each left-hand-side block contains a depthwise convolution of size \(k\times k\times k\), an expansion layer that multiplies channels by an expansion ratio, and a compression layer with \(1\times1\times1\) kernels to project back to the residual pathway. The architecture uses channel-wise GroupNorm and GELU activation [2303.09975]. Upsampling and downsampling are also performed by residual inverted-bottleneck-style blocks rather than by plain resolution-change operators, with stride-2 inserted into the first depthwise layer and a matching residual projection in the shortcut path [2303.09975].

For the Large configuration, the original MedNeXt paper specifies the stagewise depth and expansion settings as
\[
B_1=B_9=3,\qquad B_2=B_8=4,\qquad B_{3-7}=8
\]
and
\[
R_1=R_9=3,\qquad R_2=R_8=4,\qquad R_{3-7}=8.
\]
All configurations use base channels \(C=32\) [2303.09975]. The paper also describes MedNeXt-L as the “largest 62-MedNeXt-block architecture,” which is consistent with the stagewise repeated blocks together with the dedicated upsampling and downsampling MedNeXt blocks [2303.09975].

For MedNeXt-L-k5 specifically, the defining architectural change relative to the \(k3\) counterpart is that the depthwise kernels are \(5\times5\times5\) throughout the MedNeXt blocks, including the corresponding resampling blocks [2303.09975].

## 3. Compound scaling and UpKern

A central contribution of MedNeXt is compound scaling over three axes: block count \(B\), expansion ratio \(R\), and kernel size \(k\). In this design, receptive-field scaling is explicitly decoupled from width scaling, so larger kernels are not merely a by-product of wider layers but a first-class architectural variable [2303.09975].

The paper also introduces **UpKern**, a kernel-upsampling initialization strategy intended to iteratively increase kernel size without the performance saturation observed when large kernels are trained directly in data-scarce medical settings [2303.09975]. The mechanism is described as trilinearly upsampling convolution kernels of incompatible size while copying unchanged weights directly for all other layers, including normalization layers [2303.09975]. In later MedNeXt-based breast MRI work, the same idea is summarized as
\[
W_{5\times5\times5}=\mathrm{TrilinearInterpolate}(W_{3\times3\times3}),
\]
with the stated purpose of enabling a “smooth transition to a large-kernel configuration” and “stable transfer of learned features to larger kernels” [2508.01831].

This suggests that MedNeXt-L-k5 is often best understood not only as a static architecture, but also as a training trajectory: a learned \(k3\) model can serve as the source for a \(k5\) model through kernel interpolation, after which the larger-kernel network is fine-tuned [2303.09975]. In the original MedNeXt study, the flagship large-kernel evaluation uses MedNeXt-L with \(5\times5\times5\) kernels initialized using UpKern [2303.09975].

## 4. Training protocol and computational profile

In the original benchmark study, MedNeXt is implemented in **PyTorch** within the **nnUNet** framework and trained with the nnUNet schedule unchanged: **1000 epochs** and **250 batches per epoch** [2303.09975]. The optimizer is **AdamW**. The default learning rate is **0.001**, with one explicit exception: **kernel 5 in KiTS19 uses 0.0001** for stability [2303.09975]. Training and inference are performed on data resampled to **1.0 mm isotropic spacing**, with a **\(128\times128\times128\)** patch size, **batch size 2** for 3D networks, and **50% patch overlap** at inference [2303.09975]. To make large 3D models trainable, the paper explicitly uses **PyTorch AMP** and **gradient checkpointing** [2303.09975].

The Large kernel-5 configuration has a concrete computational footprint. For MedNeXt-L at kernel \(5\), the reported values are **63.0M** parameters, **564 GFLOPs**, and **735 s/train epoch** [2303.09975]. Relative to MedNeXt-L at kernel \(3\), this changes parameter count from **61.8M** to **63.0M**, GFLOPs from **500** to **564**, and training time per epoch from **267 s** to **735 s** [2303.09975]. The FLOP count is given for a single **\(128\times128\times128\)** patch, and the training-time figure corresponds to one epoch with **batch size 2** and **250 batches** [2303.09975].

These measurements make clear that the \(k5\) variant is only a modest parameter increase over \(k3\), but a substantially more expensive training configuration in wall-clock time [2303.09975].

## 5. Benchmark performance and task behavior

In the original MedNeXt evaluation, MedNeXt-L-k5 is the strongest average-performing MedNeXt variant among the reported \(S\), \(B\), \(M\), and \(L\) configurations with kernels \(3\) and \(5\) [2303.09975]. Its 5-fold cross-validation results are as follows.

| Dataset | DSC | SDC |
|---|---:|---:|
| BTCV | 84.82 | 87.85 |
| AMOS22 | 89.87 | 92.95 |
| KiTS19 | 90.71 | 87.85 |
| BraTS21 | 91.46 | 90.73 |

The average scores are **89.22 DSC** and **89.85 SDC** [2303.09975]. Compared with MedNeXt-L-k3, the \(k5\) version improves the average from **89.08** to **89.22** DSC and from **89.76** to **89.85** SDC, with gains on BTCV, AMOS22, and KiTS19 and a slight drop on BraTS21 [2303.09975]. Within the kernel-5 family, performance also rises monotonically from \(S\)-k5 through \(L\)-k5, making MedNeXt-L-k5 the top-end compound-scaled variant in that study [2303.09975].

The official test-set deployment is equally important to the model’s identity. The reported flagship submission is a **single 5-fold ensemble of MedNeXt-L using \(5\times5\times5\) kernels, initialized using UpKern**, in **3d_fullres** mode, without cascade mode [2303.09975]. The corresponding test-set DSCs are **88.76** on BTCV, **91.77** on AMOS22, **91.02** on KiTS19, and **88.01** on BraTS21 [2303.09975]. In the paper’s own framing, this large-kernel Large-scale model is the practical endpoint of the MedNeXt scaling program.

Later direct use of the exact label confirms both the utility and the limits of the configuration. In PVS segmentation, MedNeXt-L-k5 achieved a voxel-level Dice of **0.88 \(\pm\) 0.06** in white matter on **200 T2-weighted** HCP-Aging scans, but only **0.58 \(\pm\) 0.09** on the corresponding T1-weighted images, and under leave-one-site-out cross validation it reached voxel-level Dice scores of **0.38 \(\pm\) 0.16** in white matter and **0.35 \(\pm\) 0.12** in basal ganglia [2508.20256]. That paper concludes that MedNeXt-L-k5 “did not outperform the nnU-Net,” despite being competitive, which narrows claims about the universal necessity of its larger-context design [2508.20256].

## 6. Later uses, non-uses, and terminological drift

The later literature shows that MedNeXt-L-k5 functions both as an attested architecture label and as a source of ambiguity. In direct downstream use, the PVS study adapts MedNeXt-L-k5 as its main experimental MedNeXt variant, describes it as the **large** configuration with a **\(5\times5\times5\)** kernel, and uses the label explicitly throughout [2508.20256]. By contrast, several later MedNeXt papers are relevant only by analogy.

The breast MRI paper "Large Kernel MedNeXt for Breast Tumor Segmentation and Self-Normalizing Network for pCR Classification in Magnetic Resonance Images" [2508.01831] is the clearest example of near-match rather than exact match. It studies a large-kernel MedNeXt configuration obtained by moving from \(3\times3\times3\) to \(5\times5\times5\) kernels via UpKern, but the authors explicitly say: “We employed the medium variant of MedNeXt.” Their reported models are \(M^3\), \(M^5_{\text{Base}}\), and \(M^5_{\text{Focal}}\), not MedNeXt-L-k5 [2508.01831]. This is precisely the case in which “large” refers to kernel size or receptive field rather than model scale.

"Comparative Analysis of nnUNet and MedNeXt for Head and Neck Tumor Segmentation in MRI-guided Radiotherapy" [2411.14752] is relevant for a different reason: it explicitly lists a **“Large model with 5x5x5 kernel size”** among the planned MedNeXt variants, which corresponds naturally to MedNeXt-L-k5 terminology, but the paper reports that the **small model with kernel size 3** was the only MedNeXt variant that trained successfully, while the small \(k5\), large \(k3\), and large \(k5\) variants “kept collapsing after a few epochs” [2411.14752]. The paper therefore attests the attempted use of the configuration, but not a successful MedNeXt-L-k5 result.

The relationship to MedNeXt-v2 requires additional caution. "MedNeXt-v2: Scaling 3D ConvNeXts for Large-Scale Supervised Representation Learning in Medical Image Segmentation" [2512.17774] says that the **base variant** of MedNeXt-v2 “follows the large (L) configuration of the MedNeXt-v1” and adopts the **52 layer depth of the original large (L) variant of MedNeXt-v1**, but the documented MedNeXt-v2 block uses a depthwise convolution with **kernel size 3**, not 5 [2512.17774]. Thus MedNeXt-v2 inherits aspects of the Large-scale template while no true “-k5” variant is defined in that paper.

Taken together, these papers show that MedNeXt-L-k5 has a precise meaning in the original MedNeXt design space and in at least one later direct application, but that the label should not be inferred automatically from any mention of “large-kernel MedNeXt.” In the literature, one must distinguish among three cases: exact MedNeXt-L-k5 usage, Large-scale MedNeXt with unspecified kernel, and non-Large MedNeXt models whose kernels have been enlarged to \(5\times5\times5\) [2303.09975].

Source: https://www.emergentmind.com/topics/mednext-l-k5