Publications

The change from face-to-face work to teleworking caused by the pandemic has induced multiple workers to spend more time than usual in front of a computer; in addition, the sudden installation of workstations in homes means that not all of them meet the necessary characteristics for the worker to be able to position himself/herself comfortably with the correct posture in front of their computer. Furthermore, from the point of view of the medical personnel in charge of occupational risk prevention, an automated tool able to quantify the degree of incorrectness of a postural habit in a worker is needed. For this purpose, in this work, a system based on the postural detection of the worker is designed, implemented and tested, using a specialized hardware system that processes video in real time through convolutional neural networks. This system is capable of detecting the posture of the neck, shoulders and arms, providing recommendations to the worker in order to prevent possible health problems, due to poor posture. The results of the proposed system show that this video processing can be carried out in real time (up to 25 processed frames/sec) with a low power consumption (less than 10 watts) using specialized hardware, obtaining an accuracy of over 80% in terms of the pattern detected.

Diagnosis aid systems that use image analysis are currently very useful due to the large workload of health professionals involved in making diagnoses. In recent years, Convolutional Neural Networks (CNNs) have been used to help in these tasks. For this reason, multiple studies that analyze the detection precision for several diseases have been developed. However, many of these works distinguish between only two classes: healthy and with a specific disease. Based on this premise, in this work, we try to answer the questions: When training an image classification system with only two classes (healthy and sick), does this system extract the specific features of this disease, or does it only obtain the features that differentiate it from a healthy patient? Trying to answer these questions, we analyze the particular case of COVID-19 detection. Many works that classify this disease using X-ray images have been published; some of them use two classes (with and without COVID-19), while others include more classes (pneumonia, SARS, influenza, etc.). In this work, we carry out several classification studies with two classes, using test images that do not belong to those classes, in order to try to answer the previous questions. The first studies indicate problems in these two-class systems when using a third class as a test, being classified inconsistently. Deeper studies show that deep learning systems trained with two classes do not correctly extract the characteristics of pathologies, but rather differentiate the classes based on the physical characteristics of the images. After the discussion, we conclude that these two-class trained deep learning systems are not valid if there are other diseases that cause similar symptoms.

Skin blemishes can be caused by multiple events or diseases and, in some cases, it is difficult to distinguish where they come from. Therefore, there may be cases with a dangerous origin that go unnoticed or the opposite case (which can lead to overcrowding of health services). To avoid this, the use of artificial intelligence-based classifiers using images taken with mobile devices is proposed; this would help in the initial screening process and provide some information to the patient prior to their final diagnosis. To this end, this work proposes an optimization mechanism based on two phases in which a global search for the best classifiers (from among more than 150 combinations) is carried out, and, in the second phase, the best candidates are subjected to a phase of evaluation of the robustness of the system by applying the cross-validation technique. The results obtained reach 99.95% accuracy for the best case and 99.75% AUC. Comparing the developed classifier with previous works, an improvement in terms of classification rate is appreciated, as well as in the reduction of the classifier complexity, which allows our classifier to be integrated in a specific purpose system with few computational resources.

Many FPGAs vendors have recently included embedded processors in their devices, like Xilinx with ARM-Cortex A cores, together with programmable logic cells. These devices are known as Programmable System on Chip (PSoC). Their ARM cores (embedded in the processing system or PS) communicates with the programmable logic cells (PL) using ARM-standard AXI buses. In this paper we analyses the performance of exhaustive data transfers between PS and PL for a Xilinx Zynq FPGA in a co-design real scenario for Convolutional Neural Networks (CNN) accelerator, which processes, in dedicated hardware, a stream of visual information from a neuromorphic visual sensor for classification. In the PS side, a Linux operating system is running, which recollects visual events from the neuromorphic sensor into a normalized frame, and then it transfers these frames to the accelerator of multi-layered CNNs, and read results, using an AXI-DMA bus in a per-layer way. As these kind of accelerators try to process information as quick as possible, data bandwidth becomes critical and maintaining a good balanced data throughput rate requires some considerations. We present and evaluate several data partitioning techniques to improve the balance between RX and TX transfer and two different ways of transfers management: through a polling routine at the userlevel of the OS, and through a dedicated interrupt-based kernellevel driver. We demonstrate that for longer enough packets, the kernel-level driver solution gets better timing in computing a CNN classification example. Main advantage of using kernel-level driver is to have safer solutions and to have tasks scheduling in the OS to manage other important processes for our application, like frames collection from sensors and their normalization.

The use of deep learning solutions in different disciplines is increasing and their algorithms are computationally expensive in most cases. For this reason, numerous hardware accelerators have appeared to compute their operations efficiently in parallel, achieving higher performance and lower latency. These algorithms need large amounts of data to feed each of their computing layers, which makes it necessary to efficiently handle the data transfers that feed and collect the information to and from the accelerators. For the implementation of these accelerators, hybrid devices are widely used, which have an embedded computer, where an operating system can be run, and a field-programmable gate array (FPGA), where the accelerator can be deployed. In this work, we present a software API that efficiently organizes the memory, preventing reallocating data from one memory area to another, which improves the native Linux driver with a 85% speed-up and reduces the frame computing time by 28% in a real application.

Cervical cancer is caused in the vast majority of cases by the human papilloma virus (HPV) through sexual contact and requires a specific molecular-based analysis to be detected. As an HPV vaccine is available, the incidence of cervical cancer is up to ten times higher in areas without adequate healthcare resources. In recent years, liquid cytology has been used to overcome these shortcomings and perform mass screening. In addition, classifiers based on convolutional neural networks can be developed to help pathologists diagnose the disease. However, these systems always require the final verification of a pathologist to make a final diagnosis. For this reason, explainable AI techniques are required to highlight the most significant data to the healthcare professional, as it can be used to determine the confidence in the results and the areas of the image used for classification (allowing the professional to point out the areas he/she thinks are most important and cross-check them against those detected by the system in order to create incremental learning systems). In this work, a 4-phase optimization process is used to obtain a custom deep-learning classifier for distinguishing between 4 severity classes of cervical cancer with liquid-cytology images. The final classifier obtains an accuracy over 97% for 4 classes and 100% for 2 classes with execution times under 1 s (including the final report generation). Compared to previous works, the proposed classifier obtains better accuracy results with a lower computational cost.

Non-verbal communication is essential in the communication process. This means that its lack can cause misinterpretations of the message that the sender tries to transmit to the receiver. With the rise of video calls, it seems that this problem has been partially solved. However, people with cognitive disorders such as those with some kind of Autism Spectrum Disorder (ASD) are unable to interpret non-verbal communication neither live nor by video call. This work analyzes the relationship between some physiological measures (EEG, ECG, and GSR) and the affective state of the user. To do that, some public datasets are evaluated and used for a multiple Deep Learning (DL) system. Each physiological signal is pre-processed using a feature extraction process after a frequency study with the Discrete Wavelet Transform (DWT), and those coefficients are used as inputs for a single DL classifier focused on that signal. These multiple classifiers (one for each signal) are evaluated independently and their outputs are combined in order to optimize the results and obtain additional information about the most reliable signals for classifying the affective states into three levels: low, middle, and high. The full system is carefully detailed and tested, obtaining promising results (more than 95% accuracy) that demonstrate its viability.