Table 7 Monotonic relationships were observed for model I (H&E → CycleGAN) and model II (H&E → CUTGAN + DCLGAN). Respective models show direction and magnitude, derived from qualitative evaluations and HSFI, and are followed by the corresponding diagnostic interpretations.

From: VISGAB: Virtual staining-driven GAN benchmarking for optimizing skin tissue histology

Metric

Trend

Direction

Difference

Interpretation

Model I

Model II

Model I

Model II

Model I

Model II

H&E consistency

4.8 → 4.4

4.8 → 3.9

Positive

Positive

↓8.3%

↓18.8%

HSFI decreases proportionally with stain uniformity loss; structurally consistent trend

Melanin differentiation

4.7 → 4.2

4.7 → 3.6

Positive

Positive

↓10.6%

↓23.4%

HSFI reflects minor pigment differentiation degradation; aligns with color domain fidelity

Nuclear atypia

90 → 75

90 → 45

Positive

Positive

↓16.7%

↓50%

Drop in HSFI mirrors decline in nuclear feature fidelity; strong diagnostic concordance

Tissue architecture

95 → 90

95 → 55

Positive

Positive

↓5.3%

↓42.1%

HSFI tracks structural coherence; minor deviation from gold standard

Mitotic figure accuracy

85 → 60

85 → 30

Positive

Positive

↓29.4%

↓64.7%

Sharp drop in accuracy parallels steep HSFI loss; demonstrates sensitivity to fine-grained cell features

No blurring

95 → 85

95 → 55

Positive

Positive

↓10.5%

↓42.1%

Lower sharpness directly reduces HSFI; confirms structure-preservation alignment

Overstaining

3 → 10

3 → 20

Negative

Negative

↑233%

↑567%

HSFI decreases with overstaining; expected inverse relationship

Hallucinations

5 → 10

5 → 38

Negative

Negative

↑100%

↑660%

HSFI inversely tracks artifact prevalence; confirms perceptual reliability

Inter-rater agreement (Fleiss’ κ %)

94 → 85

94 → 63

Positive

Positive

↓9.6%

↓33%

HSFI declines in sync with consensus loss; supports reproducibility of diagnostic trust

Turing test success

92 → 81

92 → 50

Positive

Positive

↓12.0%

↓45.6%

HSFI aligns with human perceptual success drop; validates perceptual interpretability