Real-time estimation of the remaining surgery duration for cataract surgery using deep convolutional neural networks and long short-term memory

Table 5 Results of the RSD estimation

Data	Models	MAE (s)	MAE (%)	Difference in the MAE (s)	p-value ^*
Surgeon 001	Per-surgeon model	19.4 ± 24.9	6.4 ± 4.6	-	-
Surgeon 002	Per-surgeon model	36.4 ± 16.0	5.8 ± 2.9	Reference	-
	Pre-trained with Surgeon 1 model	94.5 ± 48.3	18.1 ± 7.6	+ 58.1 s	< 0.001
	Fine-tuned model (number for fine-tuning = 128 videos)	28.3 ± 19.0	5.2 ± 2.8	-8.1 s	< 0.001
Surgeon 003	Per-surgeon model	38.1 ± 18.5	9.3 ± 6.8	Reference	-
	Pre-trained with Surgeon 1 model	87.4 ± 39.9	17.2 ± 8.0	+ 49.3 s	< 0.001
	Fine-tuned model (number for fine-tuning = 27 videos)	30.6 ± 15.3	7.9 ± 4.0	-7.5 s	0.003

Per-surgeon model: a model without fine-tuning using per-surgeon’s data for training. Pre-trained model: a model only uses the pre-trained parameter from surgeon 1 to surgeons 2 and 3. Fine-tuned model: a model using the pre-trained parameter from surgeon 1 and fine-tuned with the training data. MAE: mean absolute error
^* t-test compared to the per-surgeon model

ISSN: 1472-6947