Date on Master's Thesis/Doctoral Dissertation

12-2014

Document Type

Doctoral Dissertation

Degree Name

Ph. D.

Department

Computer Engineering and Computer Science

Committee Chair

Frigui, Hichem

Committee Co-Chair (if applicable)

Calhoun, Aaron W.

Committee Member

Lauf, Adrian P.

Committee Member

Nasraoui, Olfa

Committee Member

Yampolskiy, Roman Vladimirovich

Subject

Cardiac resuscitation--Study and teaching--Simulation methods; Video recordings--Production and direction; Medical personnel--Training of--Audio-visual aids

Abstract

Medical simulations, where uncommon clinical situations can be replicated, have proved to provide a more comprehensive training. Simulations involve the use of patient simulators, which are lifelike mannequins. After each session, the physician must manually review and annotate the recordings and then debrief the trainees. This process can be tedious and retrieval of specific video segments should be automated. In this dissertation, we propose a machine learning based approach to detect and classify scenes that involve rhythmic activities such as Cardio-Pulmonary Resuscitation (CPR) from training video sessions simulating medical crises. This applications requires different preprocessing techniques from other video applications. In particular, most processing steps require the integration of multiple features such as motion, color and spatial and temporal constrains. The first step of our approach consists of segmenting the video into shots. This is achieved by extracting color and motion information from each frame and identifying locations where consecutive frames have different features. We propose two different methods to identify shot boundaries. The first one is based on simple thresholding while the second one uses unsupervised learning techniques. The second step of our approach consists of selecting one key frame from each shot and segmenting it into homogeneous regions. Then few regions of interest are identified for further processing. These regions are selected based on the type of motion of their pixels and their likelihood to be skin-like regions. The regions of interest are tracked and a sequence of observations that encode their motion throughout the shot is extracted. The next step of our approach uses an HMM classiffier to discriminate between regions that involve CPR actions and other regions. We experiment with both continuous and discrete HMM. Finally, to improve the accuracy of our system, we also detect faces in each key frame, track them throughout the shot, and fuse their HMM confidence with the region's confidence. To allow the user to view and analyze the video training session much more efficiently, we have also developed a graphical user interface (GUI) for CPR video scene retrieval and analysis with several desirable features. To validate our proposed approach to detect CPR scenes, we use one video simulation session recorded by the SPARC group to train the HMM classifiers and learn the system's parameters. Then, we analyze the proposed system on other video recordings. We show that our approach can identify most CPR scenes with few false alarms.

Share

COinS