Integration of a topic probability distribution into surgical phase estimation with a hidden Markov model

In this paper, we present two new methods to integrate latent Dirichlet allocation (LDA) which is a topic model for surgical workflow phase estimation with a hidden Markov model (HMM). The proposed methods are able to detect surgical phases automatically based on codebook which is built by quantizing the extracted optical flow vectors from the recorded videos of surgical processes. To detect the current phase at a given time point of an operation, some sets of training data with correct phase labels need to be learned by LDA. All documents which are actually short clips divided from the recorded videos are presented as mixtures over learned latent topics. These presentations are then quantized as observed values of a HMM. The major difference between two proposed methods is that while the first method quantizes all topic-based presentations based on k-means, the second method does this based on multivariate Gaussian mixture model. A Left to Right HMM is appropriate for this work because there is no switching the order between surgical phases.