Dual-Domain Dataset: Concepts and Applications

• Dual-Domain Dataset is defined as a resource where two related data domains are explicitly connected, supporting joint training and evaluation.
• They can represent paired measurement spaces, such as k-space and image space in MRI reconstruction, to facilitate cross-domain feature learning.
• Dual-domain datasets often employ a source–target split or synchronized acquisition to improve domain adaptation in tasks like segmentation and object detection.

A dual-domain dataset is a dataset, benchmark, or derived evaluation protocol in which learning is organized around two explicitly related domains. In current literature the term is not uniform. It may denote a multi-modality benchmark used with source–target roles, such as cardiac MR and CT in MM-WHS 2017; a paired measurement representation such as k-space and image space in fast MRI; or a synchronized sensor resource such as simultaneous dual 4D radar capture. It may also describe a method operating on an existing benchmark rather than a newly introduced dataset, which is a critical distinction in discussions of “dual-domain” resources (Li et al., 2021, Liu et al., 2022, Zhang et al., 2023).

1. Terminological scope

In the cited literature, “dual-domain” appears in several distinct senses. The common element is not a single file format or modality choice, but an explicit computational relation between two domains that are learned jointly.

Usage in the literature	Domain pair	Representative example
Multi-modality benchmark	MR and CT	MM-WHS 2017 in cardiac segmentation
Paired measurement spaces	k-space and image space	fastMRI and IXI reconstruction protocols
Native synchronized sensing	ARS548 RDI and Arbe Phoenix 4D radar	Dual Radar

A broader usage also appears when a standard dataset is converted into two coupled internal representations. In "Dataset Condensation for Time Series Classification via Dual Domain Matching" (Liu et al., 2024), the two domains are the time domain and the frequency domain; in "D2Diff: A Dual Domain Diffusion Model for Accurate Multi-Contrast MRI Synthesis" (Dayarathna et al., 18 Jun 2025), the two branches are spatial image features and DCT-derived frequency features; and in "SpecXNet: A Dual-Domain Convolutional Network for Robust Deepfake Detection" (Alam et al., 26 Sep 2025), ordinary RGB images are converted inside the network into spatial and spectral feature streams. This suggests that, in recent work, a dual-domain dataset may be native to data acquisition or may emerge from the way an existing dataset is operationalized (Liu et al., 2024, Dayarathna et al., 18 Jun 2025, Alam et al., 26 Sep 2025).

2. Formal source–target organization

A particularly explicit dual-domain formulation appears in cardiac semi-supervised domain adaptation. "Dual-Teacher++: Exploiting Intra-domain and Inter-domain Knowledge with Reliable Transfer for Cardiac Segmentation" defines three subsets: $$\mathcal{X}_{s}=\{(\mathbf{x}_i^s, y_i^s)\}_{i=1}^{m_s},$$

$$\mathcal{X}_{t}=\{(\mathbf{x}_i^t, y_i^t)\}_{i=1}^{m_t},$$

$\mathcal{X}_{u}=\{ \mathbf{x}_i^u \}_{i=1}^{m_u},$

with $m_t \ll m_u$. This is dual-domain learning in a strict sense because the model jointly uses a labeled source modality/domain and a target modality/domain that contains both limited labels and additional unlabeled data (Li et al., 2021).

The same line of work makes the source–target distinction operational rather than merely descriptive. In the earlier "Dual-Teacher" formulation, the source domain is labeled MR and the target domain is labeled and unlabeled CT; the student learns directly from scarce labeled CT, from an intra-domain EMA teacher on unlabeled CT, and from an inter-domain teacher trained on MR translated to CT-like appearance via CycleGAN (Li et al., 2020). Dual-Teacher++ generalizes this into a bidirectional MR $\rightarrow$ CT and CT $\rightarrow$ MR protocol and adds reliability control for both inter-domain and intra-domain transfer, but the underlying dataset logic remains the same: two domains, asymmetric supervision, and explicit source–target designation (Li et al., 2021).

This source–target structure is important because a dual-domain dataset is often not merely “two modalities present in one corpus.” It is a benchmark in which the two domains are assigned distinct roles in training, evaluation, or both.

3. MM-WHS 2017 as a canonical multi-modality dual-domain benchmark

For the cardiac segmentation literature in question, the relevant resource is not a new dataset but the existing MM-WHS 2017 benchmark, which becomes dual-domain by protocol. The paper states that MM-WHS 2017 provides 20 labeled CT volumes, 40 unlabeled CT volumes, 20 labeled MR volumes, and 40 unlabeled MR volumes. The same benchmark therefore supplies both domains/modalities—cardiac CT and cardiac MR—and supports bidirectional transfer: MR $\rightarrow$ CT and CT $\rightarrow$ MR (Li et al., 2021).

The split protocol is also explicit. In each direction, the labeled target set is divided by four-fold cross-validation. For MR $\rightarrow$ CT, the 20 annotated CT volumes are split into 4 folds; in each fold, 5 CT volumes form the labeled target training set $\mathcal{X}_t$, the remaining 15 CT volumes are used for testing, all 20 labeled MR volumes are used as source $$\mathcal{X}_{t}=\{(\mathbf{x}_i^t, y_i^t)\}_{i=1}^{m_t},$$0, and all 40 unlabeled CT volumes are used as unlabeled target data $$\mathcal{X}_{t}=\{(\mathbf{x}_i^t, y_i^t)\}_{i=1}^{m_t},$$1. The CT $$\mathcal{X}_{t}=\{(\mathbf{x}_i^t, y_i^t)\}_{i=1}^{m_t},$$2 MR direction follows the same rule symmetrically (Li et al., 2021).

Preprocessing is narrowly specified: all volumes are resampled with unit spacing, cropped centered at the heart region, and augmented on the fly with random affine transformations and random rotation. No additional intensity normalization details are specified in the text. Evaluation covers seven cardiac substructures: LV, RV, LA, RA, MYO, AA, and PA. The reported metrics are Dice coefficient [%] and Average Surface Distance (ASD) [voxel]; no HD is reported in the provided text (Li et al., 2021).

A common misconception is to classify MM-WHS usage here as a dataset contribution. The cited papers are method papers, not dataset papers. Their dataset contribution lies in making MM-WHS 2017 function as a dual-domain semi-supervised domain adaptation benchmark with precise source, target, labeled, and unlabeled partitions (Li et al., 2021, Li et al., 2020).

4. Dual-domain datasets as paired measurement spaces

In accelerated MRI reconstruction, “dual-domain” typically means two mathematically coupled representations of the same acquisition rather than two hospitals or two patient cohorts. "Dual-Domain Reconstruction Networks with V-Net and K-Net for fast MRI" uses the fastMRI single-coil knee dataset, which contains raw k-space data and DICOM images, and exploits both jointly. The reported split is 973 training volumes, 199 validation volumes, and 108 test volumes; each volume has roughly 36 slices. The experiments use Cartesian undersampling, 4× acceleration random undersampling masks, and a fully sampled central region occupying 8% of all k-space lines (Liu et al., 2022).

The duality here is explicit in the reconstruction pipeline. The overall network takes incomplete k-space matrices and the corresponding undersampled images obtained by IFFT. V-Net operates in the image domain on a two-channel complex image representation, while K-Net operates in k-space and uses cross-domain pooling and upsampling that pass through image space by FFT/IFFT. This is a dual-domain dataset only in the sense that fastMRI supplies the raw information needed to instantiate both domains (Liu et al., 2022).

A closely related but distinct construction appears in "Rethinking Dual-Domain Undersampled MRI reconstruction" (Gao et al., 2023). There the public IXI dataset is used as a paired multi-contrast brain MRI resource with 575 subjects carrying paired T2 and PD volumes. From each subject, 14 slices are uniformly sampled, yielding 5628 training images, 812 validation images, and 1610 test images. Images are center-cropped to $$\mathcal{X}_{t}=\{(\mathbf{x}_i^t, y_i^t)\}_{i=1}^{m_t},$$3, and reconstruction uses 1D Cartesian undersampling with acceleration from $$\mathcal{X}_{t}=\{(\mathbf{x}_i^t, y_i^t)\}_{i=1}^{m_t},$$4 to $$\mathcal{X}_{t}=\{(\mathbf{x}_i^t, y_i^t)\}_{i=1}^{m_t},$$5 and fixed

$$\mathcal{X}_{t}=\{(\mathbf{x}_i^t, y_i^t)\}_{i=1}^{m_t},$$6

The two domains are the k-space domain and the image domain, linked by $$\mathcal{X}_{t}=\{(\mathbf{x}_i^t, y_i^t)\}_{i=1}^{m_t},$$7 and $$\mathcal{X}_{t}=\{(\mathbf{x}_i^t, y_i^t)\}_{i=1}^{m_t},$$8, not two separately curated datasets (Gao et al., 2023).

This measurement-space interpretation broadens the meaning of dual-domain dataset. A dataset may qualify because each sample natively supports two physically coupled representations that are both used during training and inference.

5. Native synchronized dual-domain acquisition

A stricter, acquisition-level use of the term appears in autonomous driving. "Dual Radar: A Multi-modal Dataset with Dual 4D Radar for Autonomous Driving" introduces a dataset whose distinguishing property is simultaneous capture of two different 4D radar point clouds in the same scenes. The two radar domains are the Continental ARS548 RDI and the Arbe Phoenix. The dataset contains 151 consecutive sequences, most lasting 20 seconds, and 10,007 meticulously synchronized and annotated frames, with 103,272 annotated objects. Synchronization uses Precision Time Protocol (PTP), and the full sensor suite includes 1 camera, 1 LiDAR, and 2 4D radars (Zhang et al., 2023).

This is a dual-domain dataset in the most literal sense. The two radar streams are not redundant copies of the same sensor. They differ in range, field of view, and filtering behavior. The paper emphasizes a density–noise tradeoff: the Arbe Phoenix typically produces a much denser but noisier point cloud, whereas the ARS548 RDI is much sparser because of stronger denoising. The average point counts per frame make this contrast concrete: 11,172 for Arbe Phoenix versus 523 for ARS548 RDI, compared with 116,096 for LiDAR (Zhang et al., 2023).

The dataset is labeled with 3D bounding boxes, object class labels, and tracking IDs, and focuses experimentally on Car, Pedestrian, and Cyclist. In this setting, “dual-domain” is neither a source–target split nor a transform-generated feature pair. It is a native, synchronized, two-domain measurement resource designed precisely so that the same traffic scene can be analyzed under two radar regimes (Zhang et al., 2023).

A related but weaker form of acquisition pairing appears in multi-contrast MRI synthesis. "D2Diff" evaluates on BraTS2019 with 305 co-registered multi-contrast MRI volumes and on a healthy brain MRI dataset with 85 scans. There, however, the released data remain standard multi-contrast images; the duality is introduced by the model, which constructs a spatial branch from source contrasts and a frequency branch from multi-scale DCT features

$$\mathcal{X}_{t}=\{(\mathbf{x}_i^t, y_i^t)\}_{i=1}^{m_t},$$9

This clarifies the difference between a natively dual-domain dataset and a dataset that is rendered dual-domain by representation design (Dayarathna et al., 18 Jun 2025).

6. Boundary cases, taxonomy mismatch, and common misconceptions

Not every two-source or many-domain corpus should be called a dual-domain dataset. "M2D2: A Massively Multi-domain Language Modeling Dataset" contains 8.5B tokens, 145 domains, and 22 groups drawn from Wikipedia and Semantic Scholar/S2ORC. The paper itself stresses that this is a massively multi-domain and hierarchical resource rather than a dataset organized around one canonical domain pair (Reid et al., 2022). Similarly, "Domain2Vec" introduces TinyDA with 54 domains and about 965,619–1,000,000 images, and DomainBank with 56 domains and 339,772 images. These benchmarks are valuable because they permit many source–target pairs, but they are not explicit dual-domain datasets in the narrow sense (Peng et al., 2020).

A related misconception is to equate dual-domain with any multi-dataset training scenario. "Multi-domain semantic segmentation with overlapping labels" is concerned with the fact that different datasets often use incompatible or overlapping label taxonomies. Its solution is a universal taxonomy $\mathcal{X}_{u}=\{ \mathbf{x}_i^u \}_{i=1}^{m_u},$0 with probabilistic partial-label supervision,

$\mathcal{X}_{u}=\{ \mathbf{x}_i^u \}_{i=1}^{m_u},$1

but the problem is fundamentally taxonomy reconciliation across several domains, not the design of a two-domain dataset (Bevandić et al., 2021).

The literature therefore supports a narrower and more precise interpretation. A dual-domain dataset is best understood as one of the following: a benchmark containing two domains with explicit training roles; a resource whose samples admit two coupled physical or mathematical representations; or a synchronized acquisition setup that records two sensor domains under identical scene conditions. When a paper merely adds a second internal representation to an ordinary dataset, or derives many possible domain pairs from a larger collection, the result is dual-domain methodology rather than a native dual-domain dataset. This distinction is explicit in the way MM-WHS 2017, fastMRI, IXI, Dual Radar, M2D2, and DomainBank are described in the cited works (Li et al., 2021, Liu et al., 2022, Zhang et al., 2023, Reid et al., 2022, Peng et al., 2020).