Structural Similarity Index (SSIM)

Measures the similarity of mean intensity between the generated and reference images. This component evaluates whether the overall brightness level is preserved, using the average of pixel values. The comparison function is designed to saturate at high luminance levels, mimicking the Weber-Fechner law of human visual perception where sensitivity to absolute brightness changes diminishes as luminance increases.

• Computed as a function of mean intensities: l(x, y) = (2μxμy + C1) / (μx² + μy² + C1)
• C1 is a stabilization constant to avoid division by zero
• Sensitive to global shifts in brightness that MSE would penalize equally across all intensity ranges

Evaluates the similarity of signal contrast by comparing the standard deviation of pixel intensities. This component assesses whether the range and distribution of brightness values match between images, effectively measuring the dynamic range fidelity. A synthetic medical image with washed-out contrast will score poorly here even if structural edges align.

• Computed as a function of standard deviations: c(x, y) = (2σxσy + C2) / (σx² + σy² + C2)
• C2 is a stabilization constant
• Critical for preserving tissue differentiation in synthetic CT and MRI generation

Quantifies the correlation between the normalized pixel values of the two images. This is the defining component of SSIM, capturing the spatial interdependence of pixels that constitutes structural information. By normalizing luminance and contrast first, the structure term isolates the underlying pattern of the image—the arrangement of edges, textures, and shapes that the human visual system prioritizes.

• Uses the covariance σxy normalized by the product of standard deviations
• Formula: s(x, y) = (σxy + C3) / (σxσy + C3)
• This is what makes SSIM superior to MSE for perceptual tasks

SSIM is not computed globally over the entire image. Instead, it is applied locally using a sliding Gaussian window (typically 11x11 pixels with σ=1.5), and the results are averaged to produce a single quality score. This local approach accounts for the fact that the human visual system perceives images in localized regions and that distortions can vary spatially.

• Prevents a single large artifact from dominating the score
• Produces a SSIM map that visualizes quality variations across the image
• The final Mean SSIM (MSSIM) is the average of all local window scores

An evolution of the original SSIM that incorporates image details at multiple resolutions through iterative low-pass filtering and downsampling. Multi-Scale SSIM evaluates structural fidelity at various scales, making it more robust to viewing distance variations and better correlated with subjective human ratings than single-scale SSIM.

• Applies contrast and structure comparisons at each scale
• Luminance comparison is performed only at the coarsest scale
• Particularly valuable for evaluating super-resolution GANs and medical image reconstruction where fine detail preservation is critical

Traditional metrics like Mean Squared Error (MSE) and Peak Signal-to-Noise Ratio (PSNR) operate on per-pixel differences and treat all errors equally, regardless of their perceptual impact. SSIM addresses their fundamental flaw: two images with identical MSE can have drastically different perceived quality.

• MSE is blind to structural distortions like blurring or blocking artifacts
• SSIM values range from -1 to 1, where 1 indicates perfect structural similarity
• In medical imaging, SSIM better correlates with radiologist assessments of synthetic image diagnostic quality than PSNR