MuCED: Dual Datasets in Pathology & Music

• MuCED is an overloaded acronym referring to two datasets: one for automated cell detection in duodenal biopsy histopathology and one for direct music-to-palette generation based on emotion alignment.
• The computational pathology dataset (MuCeD) offers 55 high-resolution biopsy images with 8,608 annotated cells for multi-class cell detection, while the affective computing dataset contains 2,634 expert-validated music–palette pairs aligned through Russell’s circumplex model.
• Distinct design choices and validation protocols, such as [email protected] for pathology and cosine similarity for emotion vectors in music, underscore the importance of disambiguating MuCED in research contexts.

Searching arXiv for the two papers associated with the term "MuCED" to ground the article and disambiguate usage. Searching arXiv for ([2304.00741](/papers/2304.00741)) and ([2507.04758](/papers/2507.04758)). MuCED is an overloaded acronym in recent arXiv literature. It denotes two unrelated datasets in two different research programs: MuCeD, often also written MuCED, the Multi-class Celiac Disease Dataset for multi-class cell detection and counting in duodenal biopsy histopathology; and MuCED, the Music–Color Emotion Dataset for direct music-to-palette generation through a shared affective representation. The former is introduced as a task-specific benchmark and application domain for guided posterior regularization in computational pathology (Tyagi et al., 2023). The latter is introduced as the supervision substrate for cross-modal representation learning between music and color palettes in affective computing (Hu et al., 7 Jul 2025).

1. Terminological scope and disambiguation

In the pathology literature, MuCeD refers to a dataset of H&E-stained duodenal biopsy images designed for automated detection and counting of intraepithelial lymphocytes (IELs) and epithelial nuclei (ENs), with explicit relevance to celiac disease diagnosis. In the multimedia and affective-computing literature, MuCED refers to a dataset of music–palette pairs aligned in a shared Russell-based emotion space. These uses are semantically unrelated despite the near-identical acronym (Tyagi et al., 2023, Hu et al., 7 Jul 2025).

Variant	Domain	Core contents
MuCeD / MuCED	Computational pathology	55 duodenal biopsy images; 8,608 annotated cells
MuCED	Cross-modal affective computing	2,634 expert-validated music–palette pairs

This naming collision matters operationally. A reference to “MuCED” without expansion is ambiguous unless the surrounding context specifies either celiac histopathology or music–color emotion alignment. A plausible implication is that bibliographic, benchmarking, and dataset-indexing systems should treat the two resources as distinct entries rather than as alternate spellings of one corpus.

2. MuCeD in computational pathology

MuCeD in (Tyagi et al., 2023) is a new, task-specific dataset for multi-class cell detection and counting (MC2DC) in H&E-stained histopathology images from human duodenum biopsies, aimed at predicting / assisting diagnosis of celiac disease. Its clinical motivation is the pathologist workflow in which celiac disease diagnosis often relies on counts of IELs relative to ENs. Manual counting is described as tedious and susceptible to high inter- and intra-observer variability, so the dataset is intended to support automatic and reproducible quantification (Tyagi et al., 2023).

The dataset contains 55 high-resolution biopsy images and 8,608 annotated cells, split into 2,090 IELs and 6,518 ENs. The images are divided into five folds of 11 images each; for each fold, the protocol uses 36 images for training, 8 for validation, and 11 for testing, with results reported under 5-fold cross-validation. The class distribution is strongly imbalanced, with an approximate IEL:EN ratio of 1:3.1, and the paper reports approximately 156 cells per image when discussing density (Tyagi et al., 2023).

Acquisition and annotation are tightly controlled. The tissue is duodenum mucosa, stained with hematoxylin and eosin, imaged using an Olympus BX50 microscope with a DP26 camera at 20× optical zoom, with original image size 1920 × 2148 pixels. Annotation uses LabelMe. Experts explicitly segment the epithelial area as the region of interest, only cells inside that region are retained, and the rest of the image is masked out. The labels are bounding boxes for exactly two classes: IELs and ENs. The dataset is described as “carefully curated and validated by expert pathologists” and “curated and validated by multiple pathologists,” although no explicit inter-observer agreement statistic is reported (Tyagi et al., 2023).

The label taxonomy is clinically grounded. IELs are characterized as smaller, circular, and darker stained nuclei within the epithelium; ENs are bigger, elongated, and lighter stained epithelial nuclei. The diagnostic rule used downstream is the Q-histology ratio

$R = \frac{\#\text{IEL}}{\#\text{EN}} \times 100,$

with celiac assigned when $R \ge 25$, corresponding to the Q-histology system noted in the paper via Das et al. 2019. This makes MuCeD not merely a generic nuclei dataset, but a fine-grained counting benchmark tied to a ratio-based clinical endpoint (Tyagi et al., 2023).

3. MuCeD as a benchmark for DeGPR

Within (Tyagi et al., 2023), MuCeD is both an application domain and a validation benchmark for DeGPR, a guided posterior regularization framework for multi-class cell detection and counting. The base detection loss is written as

$$\mathcal{L}_{\text{det}} = \mathcal{L}_{\text{obj}} + \mathcal{L}_{\text{cls}} + \mathcal{L}_{\text{loc}},$$

where the components correspond to objectness, class prediction, and localization. For MuCeD, each original image is divided into 9 subimages arranged effectively as a 3×3 grid, each subimage is rescaled to 640×640, and detection is performed on these patches. For encoder training and posterior regularization, bounding-box crops from both ground truth and predictions are resized to 224×224 (Tyagi et al., 2023).

MuCeD receives several dataset-specific design choices. The ROI masking restricts inference to the clinically relevant epithelial compartment. Detection performance is reported at IoU = 0.3, i.e., mAP@0.3, rather than the more common 0.5 threshold, because the nuclei are small and dense and the boxes may be noisier. Counting is aggregated from the nine patches back to the original-image level, and is evaluated with MAE and MRE, separately for IELs and ENs (Tyagi et al., 2023).

DeGPR exploits both explicit and implicit discriminative features between the two classes. The explicit features are box size and mean intensity:

$f_S(b) = (w_R - w_L)\cdot(h_R - h_L),$

$$f_I(b) = \frac{\sum_{h=h_L}^{h_R}\sum_{w=w_L}^{w_R} I(w,h)}{f_S(b)}.$$

These encode the pathologist prior that ENs are larger and lighter, whereas IELs are smaller and darker. The implicit features are learned by a ResNet-18 encoder trained with supervised contrastive loss on cropped nuclei patches. The resulting 512-dimensional features are reduced via PCA to a 10–22 dimensional vector preserving 90\% of variance, and class-balanced sampling is used to mitigate the IEL/EN imbalance (Tyagi et al., 2023).

Posterior regularization is imposed on the distribution of class-discriminative feature differences. For each feature $f_j$ and class $c$, the framework computes the average feature within an image,

$$\mathcal{A}_{f_j}(c) = \frac{1}{|B_c|}\sum_{b\in B_c} f_j(b),$$

then the discriminative difference

$$\mathcal{D}_{f_j}(\text{IEL}, \text{EN}) = \mathcal{A}_{f_j}(\text{IEL}) - \mathcal{A}_{f_j}(\text{EN}).$$

The distribution of these discriminative vectors is modeled by Gaussian Mixture Models separately for ground truth and predictions, and the regularizer approximates the corresponding KL divergence using $10^5$ samples. For MuCeD, the overall objective is

$R \ge 25$0

with $R \ge 25$1 and a 1:1 relative weighting between explicit and implicit regularizers (Tyagi et al., 2023).

The empirical results show consistent gains across three detector families. On MuCeD, YOLOv5 improves from 0.751 to 0.787 [email protected], Faster-RCNN from 0.496 to 0.541, and EfficientDet from 0.414 to 0.425 when DeGPR is added. The minority-class counting gains are particularly marked: for YOLOv5, IEL MAE drops from 8.97 to 5.83 and IEL MRE from 42.62 to 24.19; similar relative-error reductions are reported for Faster-RCNN and EfficientDet. Under the downstream Q-histology rule, YOLOv5 + DeGPR improves accuracy from 0.7455 to 0.877 and F1-score for the celiac class from 0.774 to 0.902. The ablation study further indicates that explicit priors alone and implicit features alone are both beneficial, and that the full configuration with balanced sampling and augmentation yields the best mAP and the best overall trade-off (Tyagi et al., 2023).

These results position MuCeD as a benchmark for fine-grained, clinically structured cell counting rather than generic nuclei detection. Relative to CoNSeP and MoNuSAC, the paper frames MuCeD as more representative of a fine-grained, low-contrast, domain-specific ratio-based diagnostic task, because its two categories occupy the same epithelial compartment and differ by subtler morphological cues (Tyagi et al., 2023).

4. MuCED in cross-modal affective computing

MuCED in (Hu et al., 7 Jul 2025) is the Music–Color Emotion Dataset, introduced for direct music-to-palette generation with emotion alignment. It consists of 2,634 expert-validated music–palette pairs, with each pair aligned in a shared emotion space derived from Russell’s circumplex model. The dataset is the supervision substrate for the Music2Palette framework: each sample contains a music clip, a corresponding color palette, and emotion vectors for both modalities (Hu et al., 7 Jul 2025).

The dataset is designed to address two limitations identified in prior work. First, some music–color systems predict only a single dominant color, which loses emotional nuance and within-palette variation. Second, other methods proceed indirectly through music → text or music → images → palettes, which may degrade affective information. MuCED is therefore constructed to support direct cross-modal learning from audio to palettes while maintaining explicit emotional correspondence (Hu et al., 7 Jul 2025).

The final dataset contains 2,634 retained pairs, with an average cosine similarity of 0.76 between music and palette emotion vectors and an average expert rating of 4.36 on a 5-point Likert scale. The music pool is drawn from Emotify (400 tracks), DEAM (1,802 tracks), and PMEmo (794 tracks). Audio is stored or resampled at 22,050 Hz; most excerpts are 45 seconds long, with some variable-length cases. For model training, the audio is transformed into a 128-bin Mel spectrogram plus chroma, spectral contrast, tonal centroid, rhythm, RMS energy, and pitch, yielding a 539-dimensional feature per frame (Hu et al., 7 Jul 2025).

The palette pool initially contains 5,992 unique palettes collected from Color Hunt, Color Lovers, and Kobayashi’s Color Image Scale dataset. Each palette has 3–5 colors. Deduplication is performed by sorting palettes according to the relative order of Lightness (L), Chroma (C), and Hue (H) in CIELCh space, removing near-duplicates that differ only by order. The quantitative experiments use an 80\% / 10\% / 10\% train/validation/test split, corresponding approximately to 2,107 / 263 / 263 pairs (Hu et al., 7 Jul 2025).

MuCED is thus not a generic color-emotion resource, nor a generic music-emotion corpus. It is a paired cross-modal dataset in which the supervision target is a palette, not an image or a text prompt, and in which both modalities are embedded into the same affective coordinate system (Hu et al., 7 Jul 2025).

5. Emotion representation, construction pipeline, and model use in Music2Palette

The shared representation in MuCED is based on Russell’s circumplex model of affect. Each music clip and each palette is assigned an 8-dimensional emotion vector with components corresponding to the anchors excited, happy, content, calm, depressed, sad, afraid, and angry. For music,

$R \ge 25$2

and for palettes,

$R \ge 25$3

The components are similarity scores to the eight emotion descriptors and are used in cosine-similarity calculations rather than as explicitly normalized probabilities (Hu et al., 7 Jul 2025).

The vectors are constructed algorithmically, with human experts entering at refinement and validation stages. Music emotion vectors are obtained using a pre-trained music–text model from Doh et al. 2022, which aligns music and text embeddings via contrastive learning. Palette emotion vectors are computed using text-embedding-3-large and emotion descriptors. Candidate music–palette matches are then ranked by cosine similarity,

$R \ge 25$4

and for each music track the system selects the top 5 palettes for human refinement (Hu et al., 7 Jul 2025).

The human pipeline has two stages. In Stage 1, 20 experts, mainly from art and design, refine the candidate palettes. Each music track is assigned to three experts, who independently adjust hue, recalibrate saturation or brightness, rearrange colors, or redesign the palette if necessary. The experts are instructed to follow three criteria: emotion consistency under Russell’s model, relationship with musical features such as pitch, loudness, and rhythm versus palette brightness, saturation, and contrast, and color complexity and diversity. This yields three distinct refined palettes per music track. In Stage 2, 25 experts evaluate the pairs; each pair is rated by at least 10 experts, including at least 3 from different cultures. Pairs with average score < 3.5 are discarded, and when multiple high-quality palettes remain for one track, only the highest-rated palette is retained (Hu et al., 7 Jul 2025).

MuCED is then used directly in the Music2Palette model. The system includes an AST-based music encoder and a color decoder operating in CIELCh space. The supervision objective combines a color distance loss, a color diversity loss, and an emotion consistency loss. The color distance term uses CIEDE2000 and an optimal permutation computed by the Hungarian algorithm:

$R \ge 25$5

$R \ge 25$6

The diversity term maximizes pairwise hue separation,

$R \ge 25$7

and the emotion term aligns the generated palette with both the music and the ground-truth palette:

$R \ge 25$8

The total loss is a weighted sum of these three objectives (Hu et al., 7 Jul 2025).

The paper frames MuCED as “the first extensive cross-modal dataset for music-to-palette generation.” That claim is part of its positioning relative to music-only emotion datasets such as Emotify, DEAM, and PMEmo, music–image datasets, text–palette datasets, and prior color–emotion studies. A plausible implication is that MuCED’s distinctiveness lies not simply in multimodality, but in using palette generation rather than image retrieval or textual mediation as the target task (Hu et al., 7 Jul 2025).

6. Comparative significance, access, and limitations

The two MuCEDs occupy different methodological niches. The pathology MuCeD is a small, expert-curated biomedical detection benchmark centered on tiny, overlapping nuclei, strong class imbalance, and a ratio-based clinical endpoint. The music–color MuCED is a cross-modal affective dataset centered on emotion-aligned generative supervision between audio and color palettes. Their shared acronym conceals a sharp divergence in modality, label structure, and intended downstream use (Tyagi et al., 2023, Hu et al., 7 Jul 2025).

Their quality-control regimes also differ. MuCeD relies on expert pathologists to define the epithelial ROI and validate cell annotations, but does not report explicit numerical inter-observer agreement. Music–Color MuCED uses algorithmic emotion-vector construction followed by two-stage expert refinement and evaluation, and reports concrete retained-dataset quality indicators: 0.76 average cosine similarity and 4.36 average expert rating. This suggests two different validation philosophies: pathology-oriented spatial annotation curation on one side, and cross-modal perceptual alignment with aesthetic filtering on the other (Tyagi et al., 2023, Hu et al., 7 Jul 2025).

The access situation is likewise asymmetric. For the pathology dataset, the authors state that they release the dataset and code and provide a GitHub repository, https://github.com/dair-iitd/DeGPR. For the music–color dataset, the paper does not provide a direct download link, license details, or a file-format specification. The text permits only limited inference that audio is stored as files at 22,050 Hz, palettes are represented as 3–5 colors in CIELCh or a convertible representation, and both modalities carry 8-dimensional real-valued emotion vectors (Tyagi et al., 2023, Hu et al., 7 Jul 2025).

The principal limitations are also different. MuCeD has only 55 images, so the paper emphasizes cross-validation, careful regularization, ROI restriction, and sensitivity to the chosen IoU threshold, with [email protected] argued to be more informative than [email protected] for this setting. Music–Color MuCED, although substantially larger at 2,634 pairs, is still modest compared with large-scale multimodal corpora and may inherit source biases from Emotify, DEAM, PMEmo, online palette communities, and Kobayashi’s scale. The paper also notes possible cultural bias in color–emotion associations and calls for broader cross-cultural and multimedia evaluation (Tyagi et al., 2023, Hu et al., 7 Jul 2025).

Taken together, MuCeD/MuCED is best understood not as a single dataset but as a homonymous label applied to two specialized resources: one for computational pathology and celiac-disease-oriented cell counting, and one for affective cross-modal learning between music and color palettes. Any technical discussion of “MuCED” is therefore incomplete unless it specifies which expansion, modality, and benchmark protocol are intended.