Table 23 Impact of synthetic data proportion on real-world generalization.

From: Novel metaheuristic optimized latent diffusion framework for automated oral disease detection in public health screening

Synthetic ratio

Real test accuracy (95% CI)

Synthetic test accuracy (95% CI)

Generalization gap (95% CI)

Rare pathology sensitivity (95% CI)

False positive rate (95% CI)

Training time (hours)

Model robustness score (mean ± SD)

ANOVA F-statistic

p-value

0% (real only)

81.4% (79.8–83.0)

N/A

N/A

67.3% (64.1–70.5)

4.7% (3.9–5.5)

8.2

6.8 ± 0.4/10

–

–

25% synthetic

89.7% (88.4–91.0)

91.2% (90.1–92.3)

1.5% (0.8–2.2)

78.4% (76.1–80.7)

3.2% (2.6–3.8)

10.1

7.9 ± 0.3/10

23.47

< 0.001

50% synthetic

94.2% (93.1–95.3)

95.8% (94.9–96.7)

1.6% (0.9–2.3)

84.7% (82.8–86.6)

2.8% (2.3–3.3)

12.3

8.7 ± 0.2/10

45.82

< 0.001

75% synthetic

96.1% (95.2–97.0)

97.3% (96.6–98.0)

1.2% (0.6–1.8)

89.2% (87.6–90.8)

2.4% (1.9–2.9)

14.2

9.2 ± 0.2/10

67.94

< 0.001

100% synthetic

92.3% (91.1–93.5)

96.7% (95.9–97.5)

4.4% (3.5–5.3)

85.6% (83.5–87.7)

3.1% (2.5–3.7)

15.8

8.4 ± 0.3/10

38.12

< 0.001

Back to article page