Table 1 Quantitative evaluation results of segmentation performance for real and synthetic data

From: Addressing data scarcity in nanomaterial segmentation networks with differentiable rendering and generative modeling

Domain

Model

DSC

AP50

PQ

 

Model–Real

0.968 ± 3.16 × 104

0.737 ± 0.014

0.938 ± 5.93 × 10−4

TiO2 HIM

Model–Synth Mill et al.

0.906 ± 0.009

\(\underline{0.493\pm 0.020}\)

0.829 ± 0.015

 

Model–Synth Ours

\(\underline{0.932\pm 0.003}\)

0.393 ± 0.016

\(\underline{0.874\pm 0.005}\)

 

Model–Real

0.955 ± 9.49 × 104

0.945 ± 0.016

0.914 ± 0.002

SiO2 HIM

Model–Synth Mill et al.

0.786 ± 0.002

0.375 ± 0.004

0.659 ± 0.003

 

Model–Synth Ours

\(\underline{0.860\pm 4.86\times 1{0}^{-4}}\)

\(\underline{0.478\pm 0.011}\)

\(\underline{0.754\pm 6.21\times 1{0}^{-4}}\)

 

Model–Real

0.964 ± 0.001

0.567 ± 0.012

0.930 ± 0.001

TiO2 SEM

Model–Synth Rühle et al.

0.911 ± 0.001

0.467 ± 0.017

0.837 ± 0.002

 

Model–Synth Ours

\(\underline{0.916\pm 0.003}\)

\(\underline{0.474\pm 0.033}\)

\(\underline{0.845\pm 0.004}\)

  1. The table presents the mean and variance of test performance across three runs, measured by the Dice Similarity Coefficient (DSC), Average Precision at 50% IoU (AP50), and Panoptic Quality (PQ) for different segmentation models trained on real and synthetic datasets across various domains: TiO2 in HIM, SiO2 in HIM, and TiO2 in SEM. The “Model–Real" rows represent the averaged test performance of the real-data models. “Model–Synth Mill et al." refers to models trained on synthetic data generated by Mill et al.28. Similarly, “Model–Synth Rühle et al." refers to models trained on synthetic data generated by Rühle et al.23. “Model–Synth Ours" refers to models trained on synthetic data generated by our DiffRenderGAN approach. Bold values indicate the best scores for each metric within a domain, and underlined values highlight the top scores among synthetic models.