SwinUNETR-v2 in Breast MRI Classification

• The paper presents a mask-guided, multi-channel SwinUNETR framework for breast MRI classification, repurposing the unchanged canonical backbone for lesion detection.
• It employs a lightweight classification head that reads the deepest encoder output while bypassing the conventional segmentation decoder.
• The system integrates mask-guided preprocessing with multi-channel 2D MIP inputs and ensemble averaging to improve sensitivity and robustness on multi-center datasets.

Searching arXiv for the source paper and canonical SwinUNETR references. {"query":"(Joshi et al., 28 Aug 2025) SwinUNETR MONAI canonical SwinUNETR arXiv", "max_results": 5} I’ll look up the cited paper and the original SwinUNETR work on arXiv so the article can ground terminology and comparisons in the literature. to=arxiv_search 彩神争霸平台json {"query":"(Joshi et al., 28 Aug 2025)", "max_results": 10} SwinUNETR-v2, in the context of breast MRI classification, does not denote a second-generation revision of the SwinUNETR architecture in the source work that is often associated with the term. The system reported in "Mask-Guided Multi-Channel SwinUNETR Framework for Robust MRI Classification" is a mask-guided, multi-channel classification pipeline built around MONAI’s canonical SwinUNETR backbone, used unchanged for feature extraction, with a lightweight classification head, breast-region masking, multi-channel dynamic contrast-enhanced MRI maximum-intensity-projection inputs, and a probability-averaging ensemble. Its target task is per-breast classification into no lesion, benign lesion, or malignant lesion on the ODELIA multi-center breast MRI challenge dataset (Joshi et al., 28 Aug 2025).

1. Architectural identity and nomenclature

The central terminological point is that the reported framework is not a "v2" of SwinUNETR. The backbone is MONAI’s canonical SwinUNETR used as-is; no transformer-stage changes, no revised encoder-decoder topology, and no modifications to the Swin Transformer building blocks are introduced. The novelty is explicitly system-level: mask-guided input preparation, multi-channel 2D MIPs derived from DCE MRI, a minimal classification head, and a simple ensemble strategy (Joshi et al., 28 Aug 2025).

This distinction matters because SwinUNETR ordinarily denotes an encoder-decoder architecture with U-Net-style skip connections and hierarchical Swin Transformer stages. In the reported framework, that architectural identity is preserved. What changes is the surrounding classification pipeline. The segmentation decoder is not used during classification training; it exists only in the separate model that produces whole-breast masks. A plausible implication is that the term "SwinUNETR-v2" can become misleading when applied to this system, because it suggests an architectural successor rather than a pipeline-level variant.

2. Backbone configuration and classification head

The backbone follows the canonical MONAI SwinUNETR configuration. In practice, the encoder comprises four hierarchical Swin Transformer stages with shifted-window self-attention and patch merging between stages. The reported canonical settings are depths per stage of $[2,2,2,2]$, numbers of attention heads $[3,6,12,24]$, embedding or feature size $48$, MLP ratio $4$, and windowed attention with window size approximately $7$ using shifted windows with half-window shift. Normalization is LayerNorm inside transformer blocks, with residual connections and GELU activations, and positional treatment relies on relative position bias rather than absolute positional embeddings (Joshi et al., 28 Aug 2025).

The decoder retains the standard U-Net-like upsampling path with multi-scale fusion through skip connections from the encoder stages. However, this decoder is not operationally part of the classification pathway. Instead, classification is performed by attaching a lightweight head to the deepest encoder representation. The head consists of adaptive average pooling, flattening, and a linear layer producing three logits corresponding to $$\{\text{no lesion}, \text{benign}, \text{malignant}\}$$, followed by softmax at inference.

The concrete difference from canonical SwinUNETR is therefore narrow but important: the backbone itself is unchanged, while the model is repurposed for classification by reading out the deepest encoded feature map. This design preserves the canonical encoder while avoiding architectural intervention inside the transformer stages.

3. Mask-guided preprocessing and multi-channel DCE representation

The mask-guided component begins with a dedicated volumetric segmentation model trained on public whole-breast mask datasets: a Duke-Breast-Cancer-MRI subset with masks and Breast-Cancer-DCE-MRI with whole-breast segmentations. For mask generation, studies are reoriented to canonical space, resampled to $0.7 \times 0.7 \times 3$ mm, cropped or padded to $512 \times 512 \times 32$, and then height-cropped to $256$ via intensity-based localization. The segmentation network predicts a whole-breast mask, after which volumes are split into left and right halves (Joshi et al., 28 Aug 2025).

Mask integration is performed outside the classifier by element-wise multiplication of each input channel with the predicted breast mask. No mask concatenation and no attention gating are introduced inside the network. The stated effect is to suppress background, reduce noise, and focus feature extraction on breast parenchyma and enhancement kinetics.

The classification input is a four-channel 2D MIP representation derived from DCE MRI. The channels are: the first post-contrast phase MIP; subtraction 1, defined as first post-contrast minus pre-contrast MIP; subtraction 2, defined as second post-contrast minus pre-contrast MIP; and the last subtraction image, defined as last available post-contrast minus pre-contrast MIP. These four 2D MIPs are concatenated along the channel dimension and fed jointly to the classifier. Temporal dynamics are not modeled by an explicit temporal module; instead, they are encoded implicitly through the stacked subtraction channels and the first post-contrast phase.

The classifier therefore operates on masked, per-breast, 2D multi-channel tensors, whereas segmentation for mask generation is volumetric. This separation of volumetric localization from 2D classification is one of the defining features of the pipeline.

4. Optimization, augmentation, and ensemble construction

Classification is trained with two loss variants: standard cross entropy and inverse-frequency weighted cross entropy. The unweighted loss is reported as

$$L_{CE} = -\frac{1}{N} \sum_{i=1}^{N} \sum_{c=1}^{C} y_{i,c} \log p_{i,c}.$$

The weighted variant uses class-frequency-derived weights $[3,6,12,24]$0, yielding a second training regime with the same architecture but different class-imbalance treatment. No focal loss or Dice terms are used in classification training (Joshi et al., 28 Aug 2025).

The training regimen separates segmentation and classification. The segmentation model is trained for 100 epochs with a 5-epoch warmup, batch size 20, initial learning rate $[3,6,12,24]$1, and cosine annealing. The classification model is trained for 300 epochs with batch size 10 and the same initial learning rate and cosine schedule. Optimizer choice, weight decay, gradient clipping, and mixed precision are not specified.

Data augmentation is extensive and implemented with Albumentations for both segmentation and classification datasets. Geometric transforms include random horizontal and vertical flips, rotations, affine transforms with scaling, shearing, and translation, elastic deformations, grid and optical distortions, and random resized crops. Intensity and noise transforms include random brightness and contrast, multiplicative noise, Gaussian blur, motion blur, Gaussian noise, and coarse dropout. The paper lists the transforms but does not specify per-transform probabilities or parameter ranges.

Ensembling provides the final prediction rule. Five cross-validation folds are trained under two regimes, with and without class weights, producing ten models. Their predicted class probabilities are averaged using equal weights:

$[3,6,12,24]$2

The reported choice is $[3,6,12,24]$3 across all folds and both weight settings. The stated rationale is to balance the benign-versus-normal trade-off while improving robustness and reducing variance.

5. Dataset, labeling scheme, and empirical performance

The framework was developed for the ODELIA consortium challenge. The dataset comprises 511 studies from six European centers, acquired on scanners from multiple vendors at field strengths of 1.5 T and 3 T. Labels are assigned per breast as no lesion, benign lesion, or malignant lesion. Imaging includes T2-weighted sequences and DCE MRI with one pre-contrast and two to seven post-contrast phases. Evaluation uses five-fold cross-validation stratified by a patient-level maximum lesion label to avoid leakage (Joshi et al., 28 Aug 2025).

Per-breast processing is fundamental to the label structure. After splitting each study at 50% width into left and right halves, each breast is classified independently. Final study-level output is therefore a pair of labels, one for the left breast and one for the right.

The reported cross-validation results compare unweighted and weighted cross-entropy training by fold, using AUC, sensitivity at 90% specificity, specificity at 90% sensitivity, and a composite score defined as the mean of those three metrics.

Fold	No-weights: AUC / Sens / Spec / Score	Weights: AUC / Sens / Spec / Score
1	0.8670 / 0.6707 / 0.5915 / 0.7097	0.8072 / 0.4939 / 0.4054 / 0.5688
2	0.9078 / 0.7427 / 0.7256 / 0.7920	0.8580 / 0.5060 / 0.6280 / 0.6640
3	0.8769 / 0.6890 / 0.6311 / 0.7323	0.7578 / 0.3720 / 0.3567 / 0.4955
4	0.8887 / 0.7593 / 0.5864 / 0.7448	0.8551 / 0.7222 / 0.4599 / 0.6791
5	0.8797 / 0.7099 / 0.6481 / 0.7459	0.8116 / 0.5309 / 0.4383 / 0.5936

On the held-out test set, the ensemble across folds and weight settings achieved AUC 0.8610, sensitivity 0.6201, specificity 0.5678, and score 0.6830. The system placed second on the challenge leaderboard. Per-class and left/right-specific metrics are not reported numerically, but the confusion matrices are described as showing benign-versus-normal confusions and improved malignant detection with ensembling.

6. Reproducibility, limitations, and interpretive boundaries

Implementation details are publicly available through the reported repository: https://github.com/smriti-joshi/bcnaim-odelia-challenge.git. The reproducible workflow is described in four stages: training the mask segmentation model on the Duke and Yunnan datasets; preprocessing ODELIA studies through spacing and size standardization, cropping, left/right splitting, MIP computation, normalization, and predicted-mask application; training the classification model with and without inverse-frequency class weights using cross-validation; and averaging per-breast probabilities across folds and weight settings to produce final labels (Joshi et al., 28 Aug 2025).

Several limitations are explicit. There are no reported ablations isolating mask guidance or the multi-channel input design, and no statistical significance tests are provided. Calibration metrics such as Brier score and expected calibration error are not reported. Cross-center, vendor-specific, and field-strength-specific analyses are also absent. Likewise unreported are parameter count, FLOPs, VRAM use, training time, inference time, pruning, quantization, throughput, optimizer configuration, weight decay, gradient clipping, and mixed precision.

The listed failure modes are benign-versus-normal overlap in enhancement patterns, missed low-SNR lesions, and missed small lesions. The framework also does not contain explicit temporal modeling beyond channel stacking, and it does not apply center- or vendor-specific domain adaptation beyond resampling, normalization, masking, and augmentation. These omissions delimit what can be concluded about generalization, calibration, and mechanism.

The most accurate characterization is therefore not "SwinUNETR-v2" in an architectural sense, but a mask-guided, multi-channel SwinUNETR-based breast MRI classification framework whose contributions reside in preprocessing, input representation, and ensembling rather than in a redesigned SwinUNETR backbone.