TABLE II.
Performance of FLI-ResNet and FLI-LeNet CNN models on prediction of cells using glycolysis, OXPHOS, and glutaminolysis trained with different datasets. Values are mean +/− standard deviation for the test datasets of the fivefold cross-validation replication.
| Data type | Model | Accuracy | Precision | Recall | F1-score |
|---|---|---|---|---|---|
| Orga | FLI-ResNet | 92.6% (±2.1%) | 92.6% (±2.2%) | 93.1% (±1.8%) | 92.7% (±2.1%) |
| MEDDb | FLI-ResNet | 87.0% (±4.9%) | 89.4% (±4.1%) | 87.5% (±4.0%) | 87.4% (±4.7%) |
| MDc | FLI-ResNet | 88.0% (±5.0%) | 89.1% (±3.6%) | 89.5% (±4.1%) | 88.0% (±5.1%) |
| Orga | FLI-LeNet | 85.0% (±3.6%) | 86.7% (±2.1%) | 85.4% (±4.2%) | 85.3% (±3.6%) |
| MEDDb | FLI-LeNet | 88.1% (±1.8%) | 89.0% (±1.3%) | 88.7% (±1.9%) | 88.4% (±1.7%) |
| MDc | FLI-LeNet | 88.6% (±1.2%) | 89.3% (±1.6%) | 88.8% (±0.7%) | 88.9% (±1.2%) |
a
Org: original TPSF dataset (21 × 21 × 256).
b
MEDD: down-sampled TPSF dataset (21 × 21 × 128) with median filter.
c
MD: down-sampled TPSF dataset (21 × 21 × 128) with mean filter.