Non-invasive transabdominal fetal pulse oximetry (TFO) can provide physicians with a better metric of fetal well-being than currently-used fetal monitoring methods. There are two major challenges in this light-based measurement modality. One is in detecting the weak fetal signal at the surface of the maternal abdomen, and the other is in recovering the fetal signal from the diffuse reflectance measured, which contains a mixture of information about both maternal and fetal tissue. In this paper, we describe the TFO system we developed and evaluate its ability to recover the fetal signal in a preclinical setting. In particular, we assess its capability to measure the fetal signal over several thicknesses of maternal tissue and gauge the effectiveness of different approaches in removing the maternal influence from the mixed signal. Our results show that our TFO system, built using commodity low-cost components, can measure the highly-attenuated fetal signal through maternal tissue as thick as 5 cm. Furthermore, we determine that adaptive noise cancellation techniques are the most effective for separating the fetal signal from the maternally-influenced noise. These findings strongly support our TFO system’s ability to perform well on a variety of different pregnant body types.
Spinal cord injuries (SCI) have vast effects on day-to-day life, including the loss of sensation and control of the bladder. Since elevated bladder pressures from urine production and storage can be detrimental to renal functionality, urologists recommend performing clean-intermittent catheterization (CIC) every two to four hours throughout the day. However, limitations in mobility make the high frequency of these trips to the bathroom prohibitive. Sometimes a patient׳s bladder fills to capacity before performing CIC and eventually leaks urine, causing unnecessary embarrassment. As such, continence is the primary concern of most SCI patients. The issue in performing CIC is that it is a time-based approach, whereas urine production does not occur at a constant rate. A demand-based ’bladder almost-full’ warning system would provide more useful notifications and help SCI patients plan their bathroom trips accordingly. In this work, we explore using near-infrared light to create a discrete, wearable, non-invasive bladder state estimation system using machine learning to determine urine volume in the bladder to provide continuous monitoring for a patient throughout the day. We do this through proof-of-concept evaluation studies using Monte Carlo simulations and describe our bladder state estimation system. We also highlight some preliminary results using the system and distinguish differences in light intensity between a full and empty bladder on a volunteer.
Many patients who suffer from spinal cord injuries (SCI) also suffer from neurogenic bladder dysfunction, and lack the sensation and control of their bladder. In order to alleviate the build up of bladder pressure from urine production and promote good renal health, it is recommended to perform clean intermittent catheterization (CIC) every 2 to 4 hours throughout the day. However, since urine production is not constant, sometimes the bladder will fill with urine to capacity before the recommended CIC time causing the patient to leak, adding unnecessary embarrassment. As such, incontinence is the primary concern of many SCI patients. Sadly, there are no practical solutions available on the market that addresses this concern. In this work, we investigate using near-infrared spectroscopy to develop a wearable and non-invasive bladder volume sensing system to provide timely alerts to SCI patients based on their current bladder volume. We showcase the feasibility of such a system using an optical phantom that mimics the bladder and by performing ex vivo measurements on a pig bladder and intestines.
Current technology used for monitoring fetal well-being has been ineffective at reducing rates of harm to the fetus during the intrapartum period, yet its adoption has significantly increased the number of emergency C-sections performed. Transabdominal fetal pulse oximetry (TFO) aims to reduce the number of surgical interventions through non-invasive measurements of fetal oxygen saturation. When developing an optode for TFO, it is important to select design parameters that will maximize the measurement of the fetal signal. In this paper, we optimize the source-detector distance and wavelengths through Monte Carlo simulations using a multi-layered tissue model for various fetal depths. The results were validated by developing an optical probe with two wavelengths of light to observe pulsating arterial tissue through an optical phantom that mimics the maternal abdomen as a step towards oximetry. Our results show that 735nm and 850nm seem to be the optimal selection of peak wavelengths of light sources to obtain a stronger fetal signal for the fetal depths between 2-5 cm. Improving the signal sensitivity is approached by increasing the spacing between the source and detector, and is limited by the noise-equivalent power of the detector.
Many early stage lung cancer patients have resectable tumors, however, their cardiopulmonary function needs to be properly evaluated before they are deemed operative candidates. Such patients are typically asked to undergo standard pulmonary function tests, including cardiopulmonary exercise tests (CPET) or stair climbs. The standard tests are conducted only at selected healthcare provider locations, and are labor intensive. In addition, they are sometimes ineffective due to patient co-morbidities, such as limited mobility, which limits patient participation. To address these shortcomings, we envision that cardiopulmonary function can be evaluated in the patient’s environment using an inexpensive wearable device during routine physical activities. We present a cloud-connected mask that is fitted with CO 2 , O 2 , flow volume, and accelerometer sensors. The data collected from the device is transmitted to a cloud service, which facilitates utilization of various data mining algorithms for extraction of insights from the data. As a necessary first step toward cardiopulmonary function evaluation, we study automatic recognition of the user’s physical activity from mask sensors data via an empirical analysis of several data representation and classification algorithms. The results demonstrate accurate activity recognition using mask sensors, and underscore the potential of our approach for cardiopulmonary function evaluation.
Despite their remarkable performance in various machine intelligence tasks, the computational intensity of
Convolutional Neural Networks (CNNs) has hindered their widespread utilization in resource-constrained
embedded and IoT systems. To address this problem, we present a framework for synthesis of efficient CNN
inference software targeting mobile SoC platforms. We argue that thread granularity can substantially impact the performance and energy dissipation of the synthesized inference software, and demonstrate that
launching the maximum number of logical threads, often promoted as a guiding principle by GPGPU practitioners, does not result in an efficient implementation for mobile SoCs. We hypothesize that the runtime of a
CNN layer on a particular SoC platform can be accurately estimated as a linear function of its computational
complexity, which may seem counter-intuitive, as modern mobile SoCs utilize a plethora of heterogeneous
architectural features and dynamic resource management policies. Consequently, we develop a principled
approach and a data-driven analytical model to optimize granularity of threads during CNN software synthesis. Experimental results with several modern CNNs mapped to a commodity Android smartphone with a
Snapdragon SoC show up to 2.37X speedup in application runtime, and up to 1.9X improvement in its energy
dissipation compared to existing approaches.
Convolutional Neural Networks (CNNs) exhibit remarkable performance in various machine learning tasks. As
sensor-equipped internet of things (IoT) devices permeate into
every aspect of modern life, it is increasingly important to
run CNN inference, a computationally intensive application, on
resource constrained devices. We present a technique for fast
and energy-efficient CNN inference on mobile SoC platforms,
which are projected to be a major player in the IoT space. We
propose techniques for efficient parallelization of CNN inference
targeting mobile GPUs, and explore the underlying tradeoffs.
Experiments with running Squeezenet on three different mobile
devices confirm the effectiveness of our approach. For further
study, please refer to the project repository available on our
GitHub page: https://github.com/mtmd/Mobile_ConvNet.
Many mobile applications running on smartphones and wearable devices would potentially benefit from the accuracy and
scalability of deep CNN-based machine learning algorithms.
However, performance and energy consumption limitations
make the execution of such computationally intensive algorithms on mobile devices prohibitive. We present a GPUaccelerated library, dubbed CNNdroid, for execution of
trained deep CNNs on Android-based mobile devices. Empirical evaluations show that CNNdroid achieves up to 60X
speedup and 130X energy saving on current mobile devices.
The CNNdroid open source library is available for download
Lung cancer is the number one cause of cancer deaths. Many early stage lung cancer patients have a resectable tumor, however, their cardiopulmonary function needs to be properly evaluated before they are deemed operative candidates. Pulmonary function is assessed via spirometry and diffusion capacity. If these are below a certain threshold, cardiopulmonary exercise testing (CPET) is recommended. CPET is expensive, labor intensive, and sometimes ineffective since the patient is unable to fully participate due to comorbidities, such as limited mobility. In addition, CPET is done using a set of physical activities that may or may not be relevant to the patient’s typical activities. This paper presents steps towards developing a solution to address this gap. Specifically, we present OOCOO, a mobile mask system designed to measure oxygen and carbon dioxide levels in respiration, as well as activity levels. Unlike state of practice, oxygen, carbon dioxide, and activity data can be continuously measured over a long period of time in the patient’s environment of choice. The mask is capable of wireless data transfer to commodity smartphones. We have carried out initial work on development of an Android application to capture, analyze, and share the data with authorized entities.
We are pleased to release new results on approximation of deep Convolutional Neural Networks (CNN). We used Ristretto to approximate trained 32-bit floating point CNNs. The key results can be summarized as follows:
- 8-bit dynamic fixed point is enough to approximate three ImageNet networks.
- 32-bit multiplications can be replaced by bit-shifts for small networks.
Continue reading “On Resource-Efficient Inference using Trained Convolutional Neural Networks”