SSIMuse: Domain-Specific SSIM Extensions

• SSIMuse is a domain-specific adaptation of SSIM that applies structural similarity concepts to both symbolic music replication assessment and spatial proteomics channel synthesis.
• In symbolic music, SSIMuse-B and SSIMuse-V modify traditional SSIM components to quantify composition-level and performance-level replication, aiding in plagiarism detection and evaluation.
• In spatial proteomics, SSIMuse guides a cGAN framework through SSIM-based channel selection, enhancing generative reconstruction and scalability for high-channel imaging.

SSIMuse is a term associated with research that extends the Structural Similarity Index Measure (SSIM) beyond conventional image quality assessment. In the supplied literature, it has two distinct uses: first, as the name of a symbolic-music similarity metric for assessing data replication in MIDI-like piano-roll representations; second, as a conceptual label for an SSIM-guided conditional generative framework in multiplexed spatial proteomics. In both uses, SSIM is not merely a post hoc score: it structures representation, selection, or comparison, and is tied to questions of interpretability, fidelity, and domain-specific evaluation (Ji et al., 17 Sep 2025, Saurav et al., 2022).

1. Nomenclature and scope

The term is not used uniformly across the literature. In symbolic music, SSIMuse is explicitly introduced as “the first adaptation of the Structural Similarity Index Measure (SSIM) from images to symbolic music,” with two variants, SSIMuse-B and SSIMuse-V, aimed at composition-level and performance-level replication assessment, respectively (Ji et al., 17 Sep 2025). In multiplexed spatial proteomics, the supplied description uses SSIMuse as a conceptual system name for an SSIM-guided conditional generative adversarial pipeline that selects a subset of experimentally stained channels and synthesizes the remainder (Saurav et al., 2022). By contrast, the multimodal safety platform “MUSE” does not define an “SSIMuse” module or metric; the paper explicitly uses “MUSE” alone (Wang et al., 3 Mar 2026).

Usage	Domain	Core role of SSIM
SSIMuse-B / SSIMuse-V	Symbolic music	Replication assessment on piano-roll representations
Conceptual SSIMuse	Spatial proteomics	Channel selection and cGAN guidance
MUSE	Multimodal LLM safety	Not an SSIMuse system

A common misconception is that SSIMuse denotes a single standardized framework. The supplied sources instead support a narrower and more technical interpretation: SSIMuse designates domain-specific adaptations of SSIM, with different operational meanings in symbolic music and spatial proteomics. A second misconception is that all “SSIM-guided” methods insert SSIM directly into a training loss. In the proteomics formulation, SSIM guides channel clustering and selection rather than appearing as an explicit loss term in the cGAN objective (Saurav et al., 2022).

2. Structural-similarity basis

Both uses inherit the standard SSIM formalism. For two signals or image patches $x$ and $y$, SSIM is classically decomposed into luminance, contrast, and structure terms,

$$l(x,y) = \frac{2\mu_x\mu_y + C_1}{\mu_x^2+\mu_y^2 + C_1},$$

$$c(x,y) = \frac{2\sigma_x\sigma_y + C_2}{\sigma_x^2+\sigma_y^2 + C_2},$$

$$s(x,y) = \frac{\sigma_{xy}+C_3}{\sigma_x\sigma_y + C_3},$$

with combined form

$$\mathrm{SSIM}(x,y) = [l(x,y)]^{\alpha} [c(x,y)]^{\beta} [s(x,y)]^{\gamma},$$

and the commonly used simplified expression

$$\mathrm{SSIM}(x,y) = \frac{(2\mu_x \mu_y + C_1)(2\sigma_{xy}+C_2)} {(\mu_x^2+\mu_y^2+C_1)(\sigma_x^2+\sigma_y^2+C_2)}.$$

This formulation appears in both the proteomics and symbolic-music materials, where it is repurposed to compare channels, note-event patterns, or velocity contours rather than natural-image patches (Saurav et al., 2022, Ji et al., 17 Sep 2025).

A broader interpretive strand argues that SSIM is more fundamentally a normalized error or visibility function than a mysterious “structural” operator. In the symmetric–antisymmetric reformulation, the covariance term becomes a difference-of-variances expression, and the quantity $1-\mathrm{SSIM}$ behaves like a normalized local error energy; the square-root transformed form leads to the Dissimilarity Quotient (DQ), a noise-visibility-style measure (Larkin, 2015). This suggests that SSIMuse-like systems can be understood in two compatible ways: as direct reuses of the SSIM formula, and as domain-specific designs built around normalized, perceptually or structurally weighted error.

3. Symbolic-music SSIMuse

In symbolic music, SSIMuse is a model-agnostic, training-free similarity measure designed for replication detection in complex, polyphonic MIDI data represented as piano rolls $G \in \mathbb{R}^{T \times P}$, where $T$ is the number of time steps and $y$0 is the MIDI pitch dimension (Ji et al., 17 Sep 2025). The time quantization is fixed at 1 quarter note = 4 time steps, so that in 4/4 meter one bar corresponds to 16 time steps; the synthetic experiments use fixed-length clips of 256 time steps, i.e. 16 bars. Two piano-roll variants are used. The binary piano roll sets $y$1 at note-on positions and $y$2 otherwise, and underlies SSIMuse-B. The velocity piano roll sets $y$3 at note-on positions and $y$4 elsewhere, and underlies SSIMuse-V.

SSIMuse-B is a composition-level metric. Its luminance term reuses the standard SSIM luminance expression, but interpreted as note-density consistency. The contrast term is omitted because for binary piano rolls equal means imply equal standard deviations, making contrast redundant or ambiguous. The structure term is replaced by a Jaccard-based note-event synchronization measure on note-on positions: $y$5 This structure term is further refined in three ways. First, the piano roll is mapped to 12 pitch classes by modulo-12 aggregation,

$y$6

which provides octave invariance while preserving polyphony. Second, structure is computed over sliding time windows spanning all 12 pitch classes and aggregated with weights $y$7, with final choice $y$8. Third, one piano roll is cyclically shifted in time and pitch, and the best alignment is retained, with temporal shifts penalized by

$y$9

The result is a metric specialized to pitch–time overlap, chordal texture, rhythmic alignment, and bar-level repetition.

SSIMuse-V is a performance-level metric. It preserves the three-part SSIM structure, but each component is reinterpreted in terms of note velocities. Luminance compares average note-on velocity and thus overall dynamic level. Contrast compares the standard deviations of note-on velocities and thus dynamic dispersion. Structure is computed on a one-dimensional velocity contour obtained by collapsing polyphony at each time step to the maximum velocity among note-on events and removing silence frames. Dynamic time warping aligns the two contours, after which the standard SSIM structure term

$$l(x,y) = \frac{2\mu_x\mu_y + C_1}{\mu_x^2+\mu_y^2 + C_1},$$0

is applied to the aligned sequences. The complete performance-level score is

$$l(x,y) = \frac{2\mu_x\mu_y + C_1}{\mu_x^2+\mu_y^2 + C_1},$$1

An important misconception is that these metrics measure general stylistic similarity. The paper states a narrower objective: SSIMuse is designed explicitly for replication detection, including questions such as whether a generated clip copied at least one bar from training data, rather than for broad style matching or retrieval (Ji et al., 17 Sep 2025).

4. Evaluation in music, replication granularity, and implications

The symbolic-music study evaluates SSIMuse on Pop1K7, POP909, and POP909-melody, using a forced-replication protocol built from 16-bar clips (Ji et al., 17 Sep 2025). For each dataset, 400 reference samples and 400 mixture samples are selected, and synthetic targets are generated by inserting contiguous copied passages of 1, 2, 4, or 8 bars from a reference clip into a mixture clip. Since a 16-bar clip contains 256 time steps, these replication levels correspond to 6.25%, 12.5%, 25%, and 50% copied content. In SSIMuse-B, window size and hop size are both fixed at 16 time steps, i.e. one bar.

The paper reports that SSIMuse can reliably detect exact replication at a granularity of at least one bar. Both SSIMuse-B and SSIMuse-V increase monotonically with replication level, and Kruskal–Wallis tests show significant differences across levels with $$l(x,y) = \frac{2\mu_x\mu_y + C_1}{\mu_x^2+\mu_y^2 + C_1},$$2 (Ji et al., 17 Sep 2025). Within SSIMuse-B, the structural Jaccard component shows stronger separation than the luminance component. Within SSIMuse-V, the structural component associated with DTW-aligned dynamic contours is the most informative, while luminance changes only modestly because mean velocities can already be similar in unrelated performances.

The evaluation also positions SSIMuse against audio-embedding baselines. CLAP cosine similarity can distinguish replication levels, but requires MIDI-to-audio rendering and produces values that are close together. MERT cosine similarity fails to separate replication levels well, and the paper attributes this to insensitivity to local structural repetition. SSIMuse is therefore presented as a direct symbolic-level alternative that is fine-grained, interpretable, and separated into composition-level and performance-level components (Ji et al., 17 Sep 2025).

The music paper places these technical results in an ethical and legal context. It frames symbolic-music replication as a plagiarism and copyright concern, especially for generative systems that may memorize and reproduce training data. Because SSIMuse-B and SSIMuse-V separate score-level duplication from performance-style similarity, the metric is proposed as a basis for open evaluation, rights auditing, and monitoring of memorization in music generation. A common misconception is that replication detection is a purely legal problem external to representation design; the study instead treats representation, similarity measure, and governance as tightly linked (Ji et al., 17 Sep 2025).

5. Conceptual SSIMuse in multiplexed spatial proteomics

In multiplexed spatial proteomics, the supplied description uses SSIMuse as a conceptual name for an SSIM-guided conditional generative adversarial system that performs image-to-image synthesis of missing biomarker channels in $$l(x,y) = \frac{2\mu_x\mu_y + C_1}{\mu_x^2+\mu_y^2 + C_1},$$3-channel multiplexed images (Saurav et al., 2022). If $$l(x,y) = \frac{2\mu_x\mu_y + C_1}{\mu_x^2+\mu_y^2 + C_1},$$4 channels are experimentally collected and $$l(x,y) = \frac{2\mu_x\mu_y + C_1}{\mu_x^2+\mu_y^2 + C_1},$$5 are missing, the generator learns

$$l(x,y) = \frac{2\mu_x\mu_y + C_1}{\mu_x^2+\mu_y^2 + C_1},$$6

with $$l(x,y) = \frac{2\mu_x\mu_y + C_1}{\mu_x^2+\mu_y^2 + C_1},$$7 in the experiments. The workflow is: acquire multiplexed CODEX/PhenoCycler images; normalize and downsample each channel to $$l(x,y) = \frac{2\mu_x\mu_y + C_1}{\mu_x^2+\mu_y^2 + C_1},$$8; compute pairwise SSIM across channels; cluster channels hierarchically; select a minimal SSIM-optimal subset as conditioning input; train a U-Net-based generator in a pix2pixHD-style cGAN; and, at inference time, stain only the selected channels and synthesize the rest.

The generator is U-Net based, derived from pix2pixHD and implemented through a pix2pixHD-style codebase. The discriminator follows the pix2pixHD tradition and provides patch-wise realism scoring. The adversarial objective is the standard cGAN loss together with an $$l(x,y) = \frac{2\mu_x\mu_y + C_1}{\mu_x^2+\mu_y^2 + C_1},$$9 reconstruction term,

$$c(x,y) = \frac{2\sigma_x\sigma_y + C_2}{\sigma_x^2+\sigma_y^2 + C_2},$$0

$$c(x,y) = \frac{2\sigma_x\sigma_y + C_2}{\sigma_x^2+\sigma_y^2 + C_2},$$1

$$c(x,y) = \frac{2\sigma_x\sigma_y + C_2}{\sigma_x^2+\sigma_y^2 + C_2},$$2

A key interpretive point is that SSIM is not used as an explicit loss term in this formulation; it is used upstream as a heuristic metric for channel similarity, hierarchical clustering, and channel selection.

The channel-selection mechanism computes pairwise SSIM between channels, forms an SSIM similarity matrix, and uses hierarchical clustering to identify structurally similar groups. According to the supplied description, cluster-based selection reduces generator loss relative to random channel selection for most sampling fractions, indicating that a structurally chosen subset better covers the biology needed to reconstruct the remaining channels (Saurav et al., 2022). This SSIM-guided selection is then linked to scalability claims. Using normalized loss per prediction channel, the paper states that the architecture can scale generative image synthesis to slides with up to 100 channels, compared with prior methods limited to 11-channel data. The training data comprise 21 HuBMAP images with 29 channels from lymph node, spleen, and thymus, downsampled to $$c(x,y) = \frac{2\sigma_x\sigma_y + C_2}{\sigma_x^2+\sigma_y^2 + C_2},$$3, with an 80:20 train:validation/test split and 1000 training epochs.

The study also reports transfer to a newly generated human lung adenocarcinoma dataset. Tissue sections from 5 patients were imaged with 8 markers—CD11c, CD15, CD21, CD31, CD4, CD8, CD45, and DAPI—and a model trained on HuBMAP was applied to synthesize DAPI, CD45, and CD8 in the lung data (Saurav et al., 2022). The supplied description states that per-pixel variance for generated DAPI is comparable to HuBMAP performance, and that generated CD45 and CD8 display high morphological similarity to real stains. This is coupled to explicit ethical discussion: synthetic channels are not experimentally measured, so deployment in diagnostics would require benchmarking against gold-standard stains, transparency about virtual markers, and attention to performance drift, bias, and uncertainty.

6. Relation to broader SSIM research

SSIMuse sits within a broader research movement that treats SSIM as a reusable structural prior rather than a single fixed image-IQA formula. Related work recasts SSIM as a normalized visibility or noise-visibility function and proposes DQ as a simpler perceptually masked metric (Larkin, 2015). Other work adapts SSIM to specialized domains: MicroSSIM and MicroMS3IM address microscopy-specific issues such as high-SNR intensity scaling, detector offset, and saturation of SSIM components (Ashesh et al., 2024); DSSIM applies SSIM-style local statistics directly to floating-point simulation data after normalization and 256-bin quantization (Baker et al., 2022); and “Intensity-Sensitive Similarity Indexes for Image Quality Assessment” introduces ITW-SSIM and LISI for low-information scientific images in which important content occupies a small part of the image (Li et al., 2022).

This broader literature also shows multiple directions of methodological generalization. “Principal Component Analysis Using Structural Similarity Index for Images” replaces Euclidean reconstruction error with SSIM-based distance to define Image Structural Component Analysis (ISCA) and kernel ISCA, thereby using SSIM as the geometry of a learned image subspace (Ghojogh et al., 2019). “CSIM: A Copula-based similarity index sensitive to local changes for Image quality assessment” replaces SSIM’s luminance–contrast–structure decomposition with a patchwise Gaussian-copula representation of local pixel dependencies (Ghazouali et al., 2024). “Structured Uncertainty Similarity Score (SUSS): Learning a Probabilistic, Interpretable, Perceptual Metric Between Images” learns image-specific structured Gaussian models and scores similarity through weighted component log-probabilities, positioning itself as an interpretable probabilistic alternative to both SSIM and LPIPS (Seidler et al., 3 Dec 2025). “A Hitchhiker’s Guide to Structural Similarity” further documents that implementation choices in SSIM and MS-SSIM—window shape, stride, scaling, color handling, and pooling—materially affect performance and reproducibility (Venkataramanan et al., 2021).

This suggests that SSIMuse is best understood not as a monolithic architecture, but as part of a larger class of domain-adapted structural-similarity frameworks. In music, the adaptation shifts SSIM from images to symbolic sequences by redesigning structure and contrast terms. In proteomics, SSIM becomes a biologically motivated channel-selection heuristic for generative completion. Across these uses, the central idea remains stable: local or structured similarity statistics are treated as operational priors for selecting, reconstructing, or auditing complex data rather than as a generic fidelity score alone.