Automated detection of steps in videos of strabismus surgery using deep learning

Learning to perform strabismus surgery is an essential aspect of ophthalmologists' surgical training. Automated classification strategy for surgical steps can improve the effectiveness of training curricula and the efficient evaluation of residents' performance. To this end, we aimed to develop and validate a deep learning (DL) model for automated detecting strabismus surgery steps in the videos.

In this study, we gathered 479 strabismus surgery videos from Shanghai Children's Hospital, affiliated to Shanghai Jiao Tong University School of Medicine, spanning July 2017 to October 2021. The videos were manually cut into 3345 clips of the eight strabismus surgical steps based on the International Council of Ophthalmology's Ophthalmology Surgical Competency Assessment Rubrics (ICO-OSCAR: strabismus). The videos dataset was randomly split by eye-level into a training (60%), validation (20%) and testing dataset (20%). We evaluated two hybrid DL algorithms: a Recurrent Neural Network (RNN) based and a Transformer-based model. The evaluation metrics included: accuracy, area under the receiver operating characteristic curve, precision, recall and F1-score.

DL models identified the steps in video clips of strabismus surgery achieved macro-average AUC of 1.00 (95% CI 1.00-1.00) with Transformer-based model and 0.98 (95% CI 0.97-1.00) with RNN-based model, respectively. The Transformer-based model yielded a higher accuracy compared with RNN-based models (0.96 vs. 0.83, p < 0.001). In detecting different steps of strabismus surgery, the predictive ability of the Transformer-based model was better than that of the RNN. Precision ranged between 0.90 and 1 for the Transformer-based model and 0.75 to 0.94 for the RNN-based model. The f1-score ranged between 0.93 and 1 for the Transformer-based model and 0.78 to 0.92 for the RNN-based model.

The DL models can automate identify video steps of strabismus surgery with high accuracy and Transformer-based algorithms show excellent performance when modeling spatiotemporal features of video frames.