Leslie Casas1, Nassir Navab2, Stefanie Demirci2. 1. Computer Aided Medical Procedures, Technische Universität München, Boltzmannstr 3, 85748, Garching, Germany. leslie.casas@tum.de. 2. Computer Aided Medical Procedures, Technische Universität München, Boltzmannstr 3, 85748, Garching, Germany.
Abstract
PURPOSE: In-bed motion monitoring has become of great interest for a variety of clinical applications. Image-based approaches could be seen as a natural non-intrusive approach for this purpose; however, video devices require special challenging settings for a clinical environment. We propose to estimate the patient's posture from pressure sensors' data mapped to images. METHODS: We introduce a deep learning method to retrieve human poses from pressure sensors data. In addition, we present a second approach that is based on a hashing content-retrieval approach. RESULTS: Our results show good performance with both presented methods even in poses where the subject has minimal contact with the sensors. Moreover, we show that deep learning approaches could be used in this medical application despite the limited amount of available training data. Our ConvNet approach provides an overall posture even when the patient has less contact with the mattress surface. In addition, we show that both methods could be used in real-time patient monitoring. CONCLUSIONS: We have provided two methods to successfully perform real-time in-bed patient pose estimation, which is robust to different sizes of patient and activities. Furthermore, it can provide an overall posture even when the patient has less contact with the mattress surface.
PURPOSE: In-bed motion monitoring has become of great interest for a variety of clinical applications. Image-based approaches could be seen as a natural non-intrusive approach for this purpose; however, video devices require special challenging settings for a clinical environment. We propose to estimate the patient's posture from pressure sensors' data mapped to images. METHODS: We introduce a deep learning method to retrieve human poses from pressure sensors data. In addition, we present a second approach that is based on a hashing content-retrieval approach. RESULTS: Our results show good performance with both presented methods even in poses where the subject has minimal contact with the sensors. Moreover, we show that deep learning approaches could be used in this medical application despite the limited amount of available training data. Our ConvNet approach provides an overall posture even when the patient has less contact with the mattress surface. In addition, we show that both methods could be used in real-time patient monitoring. CONCLUSIONS: We have provided two methods to successfully perform real-time in-bed patient pose estimation, which is robust to different sizes of patient and activities. Furthermore, it can provide an overall posture even when the patient has less contact with the mattress surface.
Entities:
Keywords:
ConvNets; Deep learning; Hashing; Human pose estimation; Patient monitoring; Pressure sensors
Authors: Alexander Bigalke; Lasse Hansen; Jasper Diesel; Mattias P Heinrich Journal: Int J Comput Assist Radiol Surg Date: 2021-08-21 Impact factor: 2.924
Authors: Ariel S Kapusta; Phillip M Grice; Henry M Clever; Yash Chitalia; Daehyung Park; Charles C Kemp Journal: PLoS One Date: 2019-10-16 Impact factor: 3.240