Recovery of Surgical Workflow Without Explicit Models

Abstract

Abstract. Workflow recovery is crucial for designing context-sensitive service systems in future operating rooms. Abstract knowledge about actions which are being performed is particularly valuable in the OR. This knowledge can be used for many applications such as optimizing the workflow, recovering average workflows for guiding and evaluating training surgeons, automatic report generation and ultimately for monitoring in a context aware operating room.This paper describes a novel way for automatic recovery of the surgical workflow. Our algorithms perform this task without an implicit or explicit model of the surgery. This is achieved by the synchronization of multidimensional state vectors of signals recorded in different operations of the same type. We use an enhanced version of the dynamic time warp algorithm to calculate the temporal registration. The algorithms have been tested on 17 signals of six different surgeries of the same type. The results on this dataset are very promising because the algorithms register the steps in the surgery correctly up to seconds, which is our sampling rate. Our software visualizes the temporal registration by displaying the videos of different surgeries of the same type with varying duration precisely synchronized to each other. The synchronized videos of one surgery are either slowed down or speeded up in order to show the same steps as the ones presented in the videos of the other surgery.

Statistical modeling and recognition of surgical workflow

“…Also demonstrated on cholecystectomy, we have proposed in previous work approaches based on Dynamic Time Warping for segmenting the surgical phases of the surgery using laparoscopic tool usage [2,26]. In [27], we proposed an on-line approach using Hidden Markov Models constructed from data containing visual cues computed from the endoscopic video.…”

Surgical workflow analysis with Gaussian mixture multivariate autoregressive (GMMAR) models: a simulation study

“…Commonly, a number of presegmented signals are used to construct an ''average surgery'' (training set) that represents the sequential phases, with the corresponding timings, and this is used as a reference to segment a new operation. This approach has been applied successfully for the retrieval of 14 workflow phases of laparoscopic cholecystectomy (LC) using a set of 17 signals [5]. A boosted segmentation has been proposed by Padoy et al [6], in which DTW is combined with advanced classifiers for surgery segmentation using an adaptive distance measure based on the discriminate power of each instrument.…”

Workflow analysis and surgical phase recognition in minimally invasive surgery

“…In contrast, we are operating on low-level signals (position signal, audio and endoscopic video signals) and the identification of relevant features to segment an intervention into surgical phases. With the same focus, Ahmadi et al [2] uses a dynamic time warp algorithm for registering interventions using 17 features, which are common for each of the 14 phases of a cholecystectomy. Intraoperatively a time mapping is performed to obtain the current phase.…”