Assessment of Automated Identification of Phases in Videos of Cataract Surgery Using Machine Learning and Deep Learning Techniques

Abstract

Competence in cataract surgery is a public health necessity, and videos of cataract surgery are routinely available to educators and trainees but currently are of limited use in training. Machine learning and deep learning techniques can yield tools that efficiently segment videos of cataract surgery into constituent phases for subsequent automated skill assessment and feedback. To evaluate machine learning and deep learning algorithms for automated phase classification of manually presegmented phases in videos of cataract surgery. This was a cross-sectional study using a data set of videos from a convenience sample of 100 cataract procedures performed by faculty and trainee surgeons in an ophthalmology residency program from July 2011 to December 2017. Demographic characteristics for surgeons and patients were not captured. Ten standard labels in the procedure and 14 instruments used during surgery were manually annotated, which served as the ground truth. Five algorithms with different input data: (1) a support vector machine input with cross-sectional instrument label data; (2) a recurrent neural network (RNN) input with a time series of instrument labels; (3) a convolutional neural network (CNN) input with cross-sectional image data; (4) a CNN-RNN input with a time series of images; and (5) a CNN-RNN input with time series of images and instrument labels. Each algorithm was evaluated with 5-fold cross-validation. Accuracy, area under the receiver operating characteristic curve, sensitivity, specificity, and precision. Unweighted accuracy for the 5 algorithms ranged between 0.915 and 0.959. Area under the receiver operating characteristic curve for the 5 algorithms ranged between 0.712 and 0.773, with small differences among them. The area under the receiver operating characteristic curve for the image-only CNN-RNN (0.752) was significantly greater than that of the CNN with cross-sectional image data (0.712) (difference, -0.040; 95% CI, -0.049 to -0.033) and the CNN-RNN with images and instrument labels (0.737) (difference, 0.016; 95% CI, 0.014 to 0.018). While specificity was uniformly high for all phases with all 5 algorithms (range, 0.877 to 0.999), sensitivity ranged between 0.005 (95% CI, 0.000 to 0.015) for the support vector machine for wound closure (corneal hydration) and 0.974 (95% CI, 0.957 to 0.991) for the RNN for main incision. Precision ranged between 0.283 and 0.963. Time series modeling of instrument labels and video images using deep learning techniques may yield potentially useful tools for the automated detection of phases in cataract surgery procedures.