Table 5 Detailed search cost and segmentation performance comparison for one-shot NAS methods on BCSS, PanNuke and Zenodo Lung datasets, with U-Net based backbone

From: Large language models driven neural architecture search for universal and lightweight disease diagnosis on histopathology slide images

 

Dataset: BCSS

Dataset: PanNuke

Dataset: Zenodo Lung

Metric

Random search

Pathology-NAS

Random search

Pathology-NAS

Random search

Pathology-NAS

Iterations ↓

500

10

500

10

500

10

GPT-4 API Calls

0

10

0

10

0

10

FLOPs (G) ↓

12.63

10.58

17.72

14.33

38.45

18.52

Dice (%)

70.41 ± 0.18

74.12 ± 0.22***

88.24 ± 0.38

89.31 ± 0.44**

71.77 ± 0.46

73.94 ± 0.46***

IoU (%)

55.38 ± 0.20

59.45 ± 0.23***

80.61 ± 0.47

81.30 ± 0.45*

59.97 ± 0.39

62.05 ± 0.31***

API Cost ($)

0.00

0.14

0.00

0.15

0.00

0.16

Latency (hrs)

0.0000

0.0005

0.0000

0.0004

0.0000

0.0005

ST (GPU hrs) ↓

194.68

12.14

13.44

2.14

7.46

0.72

TT (GPU hrs) ↓

194.680

12.141

13.440

2.140

7.460

0.721

  1. Results are averaged over 5 independent runs. The table presents a comprehensive breakdown of search costs (Iterations, GPT-4 API Calls, Latency (hrs), API Cost ($), ST (GPU hrs), TT (GPU hrs)) alongside key performance metrics (FLOPs, Dice (%)). Optimal values for performance and lower values for costs are typically highlighted in bold where applicable. Statistical significance of Pathology-NAS Dice (%) and IoU (%) performance compared to Random Search (assessed by independent two-sample Welch’s t tests) is denoted by: ***p < 0.001 (very highly significant).
  2. ST Search Time, TT Total Time, TT = ST + Latency.
Back to article page