Table 2.

Drop of the weighted-averaged F1 score as a result of the perturbation of the individual feature extractor backbones

Ablative module	− ΔwAF1
Visual feature module	33.39 ± 4.46
Detection feature module	6.77 ± 2.54
Segmentation feature module	0.43 ± 0.56

The highest drop is achieved by the visual feature module, indicating that the module extracts the most important information for the prediction. The detection feature module provides a smaller impact on the model performance, while the segmentation feature module contributes only marginally to a better prediction of the surgical instruments in terms of the F1 score. The drop of the averaged F1 score over 10 folds is reported (%) with the corresponding standard deviation ( ± ).