Automated phase recognition in the cardiac catheterization laboratory with deep learning

Background An efficient workflow in the cardiac catheterization laboratory is important to guarantee quality of care and keep healthcare affordable. Automated surgical process monitoring of diagnostic coronary angiograms (CAGs) has the potential to increase workflow efficiency, reduce non-clinical workload, and optimize scheduling. The aim of this study is to automatically classify workflow phases of CAGs using deep learning models trained on a real-world dataset. Methods With an observational study we collected a dataset, composed of an integration of video recorded CAGs, clinical data, and system logging data. Every procedural phase was annotated by a medical doctor based on clinically relevant key points in time. A Scaled-YOLOv4 object detection algorithm was employed to extract features from the video images. We implemented Long short-term memory (LSTM) and LSTM-Fully convolutional network (LSTM-FCN) on the integrated time series data to classify the workflow phases. The accuracy in the classification was weighted on the duration of the phase. Results: A dataset including video, clinical and logging data of 222 CAG procedures was obtained. A weighted accuracy for phase classification of 91.3% was reached for the LSTM model, compared to 90.8% for the LSTM-FCN model. No significant difference was found after statistical comparison between both models. However, the LSTM network was 2.2 times faster in training while requiring around 40% less parameters. Conclusions This study demonstrates the potential of deploying deep learning models on a real-world integrated dataset to automatically detect CAG phases with more than 90% accuracy. No statistically significant difference was found between the performance of the two models, although the LSTM model required less training time compared to the LSTM-FCN. Thus, we suggest the employment of these models for phase recognition challenges based on composite datasets. Given the promising accuracy of these deep learning techniques, automated phase recognition could become a cornerstone of dynamic schedule optimization and increase future hospital efficiency.