ShuffleStrides Data Augmentation (SSDA) in Lung Ultrasound

• ShuffleStrides Data Augmentation (SSDA) is a clinically informed framework that augments lung ultrasound exams by permuting probe-view sequences and intra-view frames while preserving anatomical integrity.
• The method permutes the 24 valid probe-view orders and shuffles frames using prime number seeds to expand dataset diversity and enhance model training robustness.
• Integration with the permutation-invariant ZACH-ViT architecture improves sensitivity, specificity, and training efficiency, making SSDA pivotal in differentiating CPE from complex lung patterns.

ShuffleStrides Data Augmentation (SSDA) is a structured, clinically informed data augmentation framework developed for lung ultrasound (LUS) imaging, designed to expand dataset diversity and improve generalization for deep learning models tasked with classifying cardiogenic pulmonary oedema (CPE) against heterogeneous non-cardiogenic and healthy lung presentations. SSDA operates by permuting both probe-view sequences and intra-view frame orders, with all permutations constrained to respect anatomical and diagnostic integrity. Deployed with ZACH-ViT—a permutation-invariant vision transformer—SSDA drives significant advances in diagnostic accuracy, training robustness, and clinical deployment efficiency (Angelakis et al., 20 Oct 2025).

1. Concept and Rationale

ShuffleStrides Data Augmentation is motivated by the substantial heterogeneity in LUS data, especially among non-cardiogenic presentations (e.g., NCIP/ARDS-like, interstitial lung disease, and structurally normal lungs) that exhibit visually overlapping features such as B-lines and pleural artefacts. Standard augmentation methods do not account for the clinical structure intrinsic to LUS exams, where a typical dataset comprises videos from four standard thoracic probe positions. SSDA explicitly models this structure by generating all valid permutations of probe-view orderings and, in its extended form, shuffling the order of frames within each probe-view. This generates a diverse yet clinically coherent training set, enabling models to better learn permutation-invariant features suitable for unordered medical image sequences.

2. Methodological Implementation

SSDA consists of two sequential augmentation steps:

2.1 Permuting Probe-View Sequences

Given four probe-views (transducer positions), SSDA enumerates all $4! = 24$ valid orderings. For each LUS exam, the respective four videos are vertically concatenated in permuted order, yielding a stride image. This preserves the spatial and anatomical content of each probe-view:

$$\mathbf{X}_{\text{VIS}} \leftarrow \text{Concat}_{\text{vertical}}\left(\text{Resize}(V_{\pi(1)}), \dots, \text{Resize}(V_{\pi(4)})\right)$$

where $V_{\pi(i)}$ denotes the $i$-th probe-view under permutation $\pi$. The augmentation is strictly structured: only the sequence order changes, never the internal content of the views.

2.2 Permuting Frame Orders within Each Probe-View (Extended SSDA)

The 0_2-SSDA variant further randomizes the order of frames within each probe-view using a set of prime-number seeds for reproducibility and diversity (e.g., 2, 3, 5, ..., 29). For every seed, frames in each video are shuffled, and the aforementioned stride permutation step is repeated, multiplying dataset diversity up to twelvefold over the baseline. Each new composite image represents a plausible anatomical exam while scrambling temporal cues that should not influence the classification task.

Algorithm (from (Angelakis et al., 20 Oct 2025)):

def SSDA_augment(video_set, seeds=[2,3,5,7,11,13,17,19,23,29]): augmented = [] for perm in permutations([0,1,2,3]): stride = np.concatenate([resize(video_set[i]) for i in perm], axis=0) augmented.append(stride) for seed in seeds: shuffled_videos = [shuffle_frames(video, seed=seed) for video in video_set] for perm in permutations([0,1,2,3]): stride = np.concatenate([resize(shuffled_videos[i]) for i in perm], axis=0) augmented.append(stride) return augmented

3. Integration with ZACH-ViT Architecture

SSDA is tightly coupled with the Zero-token Adaptive Compact Hierarchical Vision Transformer (ZACH-ViT) architecture (Angelakis et al., 20 Oct 2025), which eschews positional embeddings and [CLS] tokens to achieve full permutation invariance—an architectural feature aligned with the unordered, permuted nature of SSDA-augmented data. The model accepts composite stride images as input, learning representations that capture clinically relevant features irrespective of probe-view or frame order. This alignment between augmentation and model design is a distinguishing attribute, contributing to improvements in classification robustness.

4. Empirical Performance and Diagnostic Impact

In experimental evaluations, SSDA significantly enhances model performance and stability:

Model	Augmentation	Validation ROC-AUC	Test ROC-AUC	Sensitivity	Specificity
ZACH-ViT	SSDA (0-SSDA)	0.87	~0.79–0.80	~0.60	~0.91
Baselines	Trivial, None	≤0.50	≤0.50	0.00	1.00

SSDA prevents collapse to degenerate solutions (trivial classifiers with zero sensitivity) by compelling the network to learn permutation-invariant, discriminative representations. The augmentation enables ZACH-ViT to differentiate CPE from complex non-cardiogenic patterns, achieving the highest ROC-AUC and balanced sensitivity/specificity among nine state-of-the-art baselines.

5. Training and Deployment Efficiency

By expanding the effective training set through clinically plausible permutations, SSDA accelerates network convergence and enhances generalization. ZACH-ViT, with only 0.25 million parameters, trains 1.35 times faster and with 2.5 times fewer parameters than Minimal ViT (0.62M parameters), supporting deployment in real-time critical care scenarios. The data augmentation strategy directly contributes to the model's resource efficiency and clinical applicability.

6. Clinical and Research Utility

SSDA’s anatomical validity ensures that generated strided images remain clinically interpretable, making it suitable for LUS-based triage and diagnostic support. Its principles may generalize to other structured medical imaging tasks where multiple views and frame sequences form the core of diagnostic reasoning, provided anatomical and diagnostic constraints are observed during permutation.

7. Context and Distinction from Related Augmentation Strategies

Unlike generic augmentation techniques—such as rotation, mirroring, or pixel-level transformations—SSDA specifically exploits domain structure by permuting probe-view sequences and intra-view frame orders under the constraint of anatomical realism. The approach contrasts with augmentation techniques that operate at the patch or image level without regard for clinical workflow, demonstrating superior results in small-data, heterogeneous medical imaging environments.

A plausible implication is that the synergy between SSDA and permutation-invariant model architectures represents a promising direction in data-efficient medical AI, especially where clinical practice prescribes specific acquisition protocols and view arrangements. SSDA exemplifies the targeted use of augmentation to improve deep model performance while constraining synthetic variation to preserve critical diagnostic features.