Enhanced sequential picking strategy with hyperparameter optimization for efficient surgical phase recognition

Accurate surgical phase recognition from surgical videos is critical for surgical training and evaluation. However, traditional models like Transformers face limitations in memory and computational efficiency, particularly as input sequence length increases. We introduce an enhanced Sequential Picking Strategy with hyperparameter optimization within the SPT-SVNet model, which integrates spatial and temporal embeddings. Through hyperparameter search, we apply different sequential picking strategies by optimizing these hyperparameters, enabling the model to selectively extract crucial features from temporal embeddings and reduce computational demands to varying degrees. Optimizing these hyperparameters led to significant performance improvements, reducing memory usage by 8.6% and training time by 16.2%, while maintaining or even enhancing model accuracy. We evaluated our approach using the SPT-SVNet model. Evaluations on the Cholec80 dataset demonstrated remarkable results, with our method achieving an accuracy of 90.5% and a precision of 92.4%, surpassing existing models.