Table 2.
Model performance on a cell-type-specific level
| Cell type | Model performance accuracy (%) |
Discard tradeoff |
|||||
|---|---|---|---|---|---|---|---|
| Deep neural network (DNN) with only handcrafted features | Multilabel k nearest neighbours with hybrid features | Random forest classifier with hybrid features | Support vector machine with hybrid features | HBNet (std. dev. Across folds) | HBNet-DHC | HBNet—DHC percentage discarded | |
| Spermatogonia (0.11) | 85.9 | 83.8 | 80.2 | 81.2 | 85.7 (0.24) | 99.4 | 37.2% |
| Preleptotene spermatocytes (0.11) | 74.8 | 85.5 | 71.9 | 73.1 | 84.9 (0.38) | 99.2 | 37.2% |
| Pachytene spermatocytes (0.22) | 69.9 | 82.4 | 72.1 | 73.3 | 82.6 (0.24) | 99.2 | 31.7% |
| Round/early spermatids (0.78) | 68.1 | 79.0 | 72.8 | 74.3 | 80.5 (0.55) | 83.5 | 39.1% |
| Elongated/late spermatids (0.22) | 77.4 | 79.8 | 76.9 | 76.7 | 85.2 (0.36) | 98.7 | 30.1% |
| Sertoli cells (1.00E-10) | 74.1 | 86.3 | 65.9 | 57.6 | 92.2 (0.16) | 92.2 | 0.0% |
| Leydig cells (0.11) | 80.4 | 81.6 | 80.3 | 76.2 | 85.7 (0.34) | 99.2 | 38.3% |
| Peritubular cells (1.00E-10) | 84.3 | 95.9 | 67.4 | 67.9 | 98.6 (0.09) | 98.7 | 1.3% |
The % accuracy for predicting the labels for each cell type is shown for standard deep neural network (DNN) with only handcrafted features, three standard classification approaches including our hybrid features (K-nearest neighbors, random forest, and support vector machines), our hybrid Bayesian neural network (HBNet), and DHC-thresholded HBNet (HBNet—DHC) along with the percentage of discarded images based on low DHC confidence. The standard deviation (std dev.) between each cross-validation fold is included for HBNet to indicate sampling variance.