The need for a better integration of new Computer-Assisted-Surgical systems in the Operating Room (OR) has been recently emphasized. One necessity to achieve this objective is to retrieve data from the OR with different sensors, then to derive models from these data for creating Surgical Process Models (SPMs). Recently, the use of videos from cameras in the OR has demonstrated its efficiency for advancing the creation of situation-aware CAS systems. The purpose of this thesis was to present a new method for the automatic detection of high-level (i.e. surgical phases) and low-level surgical tasks (i.e. surgical activities) from microscope video images only. The first step consisted in the detection of high-level surgical tasks. The idea was to combine state-of-the-art computer vision techniques with time series analysis. Image-based classifiers were implemented for extracting visual cues, therefore characterizing each frame of the video, and time-series algorithms were then applied to model time-varying data. The second step consisted in the detection of low-level surgical tasks. Information concerning surgical tools and anatomical structures were detected through an image-based approach and combined with the information of the current phase within a knowledge-based recognition system. Validated on neurosurgical and eye procedures, we obtained recognition rates of around 94% for the recognition of high-level tasks and 64% for low-level tasks. These recognition frameworks might be helpful for automatic post-operative report generation, learning/teaching purposes, and for future context-aware surgical systems.
