MedNeXt V2 Ensemble Segmentation

• MedNeXt V2 Ensemble is a deep learning segmentation framework integrating advanced MedNeXt V2 blocks, kernel expansion via UpKern, and ensembling for improved 3D medical imaging.
• The architecture employs a two-stage training regimen with small-to-large kernel fine-tuning and deep supervision to enhance performance on brain and breast tumor segmentation tasks.
• Systematic ensembling and probabilistic aggregation techniques yield consistent gains in segmentation accuracy, especially in low-resource, data-constrained clinical environments.

MedNeXt V2 Ensemble models represent a class of deep learning segmentation architectures distinguished by the use of advanced convolutional blocks (MedNeXt V2), systematic model ensembling, and application-driven training protocols for robust 3D medical image segmentation under challenging clinical and distributional settings. These ensembles have achieved leading performance on brain and breast tumor segmentation tasks, especially in low-resource or data-constrained environments, by leveraging architectural innovations, multi-scale supervision, and probabilistic aggregation.

1. MedNeXt V2 Block Architecture

MedNeXt V2 is an evolution of the original MedNeXt design, retaining ConvNeXt-inspired depthwise convolutions, channel expansion/projection, and residual skip pathways, while further introducing advanced normalization and supervision schemes (Jaheen et al., 31 Jul 2025).

Core Block Structure

• Each block consists of a pointwise (1×1×1) convolution projecting $C_{\mathrm{in}}$ input channels to $C_{\mathrm{mid}}$ (typically $2\times C_{\mathrm{in}}$).
• This is followed by a depthwise cubic convolution (typically $3\times3\times3$ or expanded to $5\times5\times5$), layer normalization (LN), and a GELU nonlinearity.
• A final pointwise projection returns to $C_{\mathrm{out}}$ (often $C_{\mathrm{in}}$), preserving dimensionality.
• Residual connections are employed throughout, and deep supervision is often imposed at decoder stages.

MedNeXt V2 blocks omit attention or squeeze–excitation modules in leading instantiations such as EMedNeXt (Jaheen et al., 31 Jul 2025). Instead, performance gains arise from increased receptive field, architectural depth, and sophisticated normalization (LayerScale, GroupNorm, LN).

2. Kernel Expansion via UpKern Algorithm

Musah et al. (Musah, 3 Aug 2025) introduced a two-stage training and kernel expansion protocol enabling stable growth of cubic convolutional filters without losing learned features or incurring instabilities.

UpKern Weight Transfer

• Given pre-trained filters $$W^3\in\mathbb{R}^{C_{\mathrm{mid}}\times C_{\mathrm{mid}}\times 3\times3\times3}$$, UpKern transfers to $W^5$ via trilinear interpolation:

$$W^5_{i,j,\alpha,\beta,\gamma} = \sum_{u,v,w=1}^{3} W^3_{i,j,u,v,w} \;\mathcal{T}(\alpha-\tfrac{u-2}{2},\,\beta-\tfrac{v-2}{2},\,\gamma-\tfrac{w-2}{2})$$

where $\mathcal{T}$ denotes the trilinear kernel.

• Filter $L_2$-norm is preserved to avoid training instability.
• This initialization allows transition from $3^3$ to $5^3$ kernels (“large-kernel fine-tuning”) while benefiting from robust features acquired during “small-kernel pretraining.”

3. Ensemble Protocols and Aggregation

Ensembling in MedNeXt V2 systems consistently leverages late-fusion at the softmax/probability level, following a uniform weighting scheme.

Model Selection and Aggregation

• Ensembles typically consist of multiple cross-validation checkpoints (“fold models”) and/or models fine-tuned with different objectives (e.g., base vs. focal-loss variants).
• Each model generates a voxel-wise softmax probability map; final predictions are computed by averaging these maps across all ensemble members:

$$P_{\mathrm{ens}}(x) = \frac{1}{N} \sum_{m=1}^N P_m(x)$$

• For binary segmentation, thresholding at 0.5 produces the final mask; for multiclass, analogous procedures apply per class.
• For example, in breast tumor segmentation, the $M^5_{\mathrm{Base}}$ and $M^5_{\mathrm{Focal}}$ models are ensembled by direct averaging of their probability maps (Musah, 3 Aug 2025). In EMedNeXt, a five-model ensemble averages the probabilities for each tumor subregion/class (Jaheen et al., 31 Jul 2025).
• Post-processing steps (e.g., topology-aware connected-component filtering, mask hierarchy enforcement) are frequently applied to enhance anatomical plausibility.

4. Training Regimens and Deep Supervision

MedNeXt V2 ensembles employ multi-stage training procedures, deep multi-scale supervision, and aggressive data augmentation to improve convergence and generalization, particularly for uncommon or low-quality imaging domains.

Two-Stage Training (Kernel Expansion)

• Stage I: small-kernel pretraining of all fold models (e.g., $3\times3\times3$ convolutions for 250 epochs) with linear Dice–cross-entropy loss and deep supervision at intermediate decoder levels (Musah, 3 Aug 2025).
• Stage II: large-kernel fine-tuning (e.g., $5\times5\times5$ convolutions, 250 epochs) with transfer initialization via UpKern. Stage II can branch: one model uses standard loss, another adds Focal loss to penalize small false negatives.
• Only the top fold(s) from Stage I are advanced to Stage II.

Deep Supervision

• Multi-layer outputs at different spatial resolutions are each equipped with their own segmentation heads.
• The total loss aggregates supervision from full- and lower-resolution outputs, using exponentially decaying weights (e.g., $w_i = 2^{-i}$ for levels $i=0,...,3$) (Jaheen et al., 31 Jul 2025).
• Loss functions typically combine Dice, Focal, and boundary penalties:

$$\mathcal{L}_{\rm seg}(p,g) = \mathcal{L}_{\rm Dice\text{-}Focal}(p,g) + \alpha \|\nabla_{\rm Sobel} p - \nabla_{\rm Sobel}g\|_2^2$$

Data Augmentation and Normalization

• On-the-fly geometric (flips, rotations, affine), intensity (scaling, shifting), and resampling are standard.
• For challenging cohorts (e.g., Sub-Saharan Africa MRIs), normalization (z-score over nonzero voxels), clipping, and foreground cropping are used (Jaheen et al., 31 Jul 2025).

5. Application to Tumor Segmentation: Quantitative Results

MedNeXt V2 ensembles have been evaluated on multiple large-scale tumor segmentation challenges. Performance improvements over MedNeXt V1 and single-model V2 systems have been consistently observed, and the architectural generality allows deployment on disparate clinical tasks.

Dataset & Task	Method/Ensemble	Mean Dice	Boundary/HD Metric
MAMA-MIA (breast DCE-MRI)	Final 2-model V2 ($5^3$) ensemble	0.67	NormHD 0.24
BraTS-Lighthouse SSA (brain MRI)	Ensemble (5× FT B3 V2) + post-processing	0.897	NSD₀.₅ 0.541, NSD₁.₀ ≈0.84
BraTS 2024 SSA (brain MRI) [V1]	B-only (5-fold) ensemble	0.896	HD95 14.7 mm
BraTS 2024 PEDs (pediatric brain MRI) [V1]	B-only (5-fold) ensemble, lr=0.0005	0.830	HD95 37.5 mm

• Compared to V1, V2 models with deep supervision and fine-tuning produced +0.045 mean Dice and +0.132 mean NSD₀.₅; ensembling yielded an additional +0.019 Dice (Jaheen et al., 31 Jul 2025).
• For breast tumor segmentation, upscaling kernel size via UpKern gained +0.02 Dice; ensembling base and focal-loss models further boosted Dice by +0.01 and reduced NormHD by 0.05 (Musah, 3 Aug 2025).

6. Clinical and Computational Considerations

Several implementation details are critical for MedNeXt V2 ensembles in clinical settings:

• Patch-based, overlapping inference manages GPU memory constraints for large 3D volumes.
• All inference pipelines employ sliding-window strategies, with local and global probability map averaging, and memory-rotation to accommodate ensemble members (Jaheen et al., 31 Jul 2025).
• Topology-aware postprocessing (connected-component pruning, hierarchical mask enforcement, merging strategies) is systematically employed across implementations to achieve anatomical coherence, especially under low-quality imaging and class imbalance.

7. Future Directions and Implications

The demonstrated utility of MedNeXt V2 ensembles in multiple clinical segmentation benchmarks highlights the value of architectural modularity, ensemble learning, and kernel-expansion protocols.

This suggests future improvements may arise through further advances in large-kernel training stability, learned aggregation (non-uniform weighting, stacking), explicit topological constraints, or integration with radiomics and clinical data for downstream predictive modeling (Musah, 3 Aug 2025). Model robustness under distribution shift and image degradation, as in SSA and pediatric settings, remains an active research focus (Jaheen et al., 31 Jul 2025, Hashmi et al., 2024).

---

References:

• [Large Kernel MedNeXt for Breast Tumor Segmentation and Self-Normalizing Network for pCR Classification in Magnetic Resonance Images, (Musah, 3 Aug 2025)]
• [EMedNeXt: An Enhanced Brain Tumor Segmentation Framework for Sub-Saharan Africa using MedNeXt V2 with Deep Supervision, (Jaheen et al., 31 Jul 2025)]
• [Optimizing Brain Tumor Segmentation with MedNeXt: BraTS 2024 SSA and Pediatrics, (Hashmi et al., 2024)]