Elsevier

Ad Hoc Networks

Volume 80, November 2018, Pages 95-103
Ad Hoc Networks

DDA: A deep neural network-based cognitive system for IoT-aided dermatosis discrimination

https://doi.org/10.1016/j.adhoc.2018.07.014Get rights and content

Abstract

The rapid development of the Internet of Things (IoT) and cognitive cyber-physical systems (CPS) has made people's daily lives more intelligent. Additionally, emerging technologies, such as wearable devices and machine learning, have demonstrated the potential for acquiring and processing large amounts of data from the physical world. In the medical field, effectively utilizing the collected medical data and providing more intelligent systems for doctors and patients to assist in diagnoses have also become important research topics. This paper presents a deep neural network-based cognitive system named DDA (dermatosis discrimination assistant) for classifying the dermatosis images generated by confocal laser scanning microscopes. Considering the lack of labels, we increase the labeled data automatically using an incremental model based on a small amount of labeled data and propose a disease discrimination model to distinguish and diagnose the categories of the disease images. In this system, the diagnoses of seborrheic keratosis (SK) and flat wart (FW) are used as examples, and experiments are conducted using the proposed models. Experimental results show that this system performs almost as well as individual dermatologists and can identify and diagnose other common dermatoses.

Introduction

The Internet of Things (IoT) and big data have been two most influential technologies of this era [1], [2]. As breakthroughs in emerging technologies, the IoT, cloud computing and data analytics provide the potential to acquire and process large amounts of data from the physical world [3], thus making it possible to acquire, process and communicate data in intelligent applications [4]. Currently, these representative technologies have been applied in many fields, such as intelligent transportation systems, smart homes, smart buildings and smart cities [5], [6]. Meanwhile, in the medical field, many scholars think that knowledge-based medical systems are a potential research topic.

Due to the development of IoT-based medical equipment, a large amount of medical data is constantly produced. Effectively analyzing and utilizing these data to provide more intelligent diagnostic tools to doctors and patients have become key issues. Recently, the deep convolutional neural network (CNN) has shown the potential to identify complex features of objects [7], and the prospect of handling medical big data has led to an interest in using deep learning to create professional medical robots with comprehensive abilities and high diagnostic accuracy [8], [9].

In the medical field, the diagnoses of dermatoses or skin diseases mainly rely on doctors' visual observations and subjective experience and lack a scientific means of quantification [10]. For example, seborrheic keratosis (SK) and flat wart (FW) are two common skin diseases. Both often occur on the face, the back of the hand and the arm. These diseases present multiple lesions and affect the appearance [11]. Thus, these diseases may distress patients and seriously affect their physical and mental health. The clinical manifestation of SK is light brown macular or flat papules, and it is characterized as a smooth or slightly raised papilloma on the skin surface. Conversely, FW is characterized by a flat papule of unequal size with a slight uplift and smooth surface, which is rounded, elliptical or polygonal. It is difficult to distinguish the two diseases because their clinical manifestations are similar. Therefore, the accurate distinction between SK and FW is highly important for timely and effective treatment [12].

To better assist doctors in diagnosing the two common skin diseases, we propose a cognitive system named DDA (dermatosis discrimination assistant) for classifying medical images based on a deep convolutional neural network. The DDA can provide diagnostic advice to doctors. In general, the training in the deep convolutional neural network requires large-scale data support. However, in the medical field, we inevitably encounter the problem of small data size and missing data labels. To address this problem, this paper uses a small amount of data labeled by doctors to automatically increase the labeled data and propose a discrimination model to distinguish and diagnose the category of the disease. Our dataset is constructed of images labeled by dermatology experts, unlabeled images and test images. The number of labeled images is small, and we use unlabeled images to increase the feature image dataset. We train the CNN using labeled images and the increased feature image dataset, which accurately identifies SK and FW.

The DDA is a cognitive system for IoT-aided dermatosis discrimination. On the one hand, in DDA, the CNN possesses the cognitive ability by learning the pathological knowledge from existing dermatological images and gives advice similar to that of a doctor using disease discrimination models. On the other hand, in practice, we can deploy the disease discrimination models on various IoT-based mobile terminals, such as mobile robots, lightweight equipments and sensors.

The remainder of this paper is organized as follows. Section 2 describes the related works. Section 3 describes the system model and key technologies, including the convolutional neural network, system work process, the dataset labeling incremental model and the performance evaluation method. Section 4 provides the experimental results. Section 5 presents the study's conclusions and outlines future work.

Section snippets

Related works

The skin is a natural barrier of the body and the first line of defense against external stimuli, such as ultraviolet light, detergents, mechanical friction and insect bites. Therefore, the incidence of skin diseases caused by environmental changes is rising. According to statistics, in the past several decades, cancer caused by skin diseases has appeared, and its serious consequences are self-evident. The traditional diagnosis of skin diseases is mostly based on clinical experience, which may

Convolutional neural network

The DDA uses CNN for dermatosis discrimination. We exploit the GoogleNet Inception v3 [7] model architecture, which is pre-trained using 1.28 million images (1000 object categories) from the ImageNet Large Scale Visual Recognition Challenge of 2014 [30], and we train it on our dataset. Fig. 1 shows the dermatosis discrimination using GoogleNet Inception v3.

The GoogleNet Inception v3 model has 46 layers. The model includes 11 Inception modules. In Fig. 1, the inception constructed in the

Incremental dataset labeling experiment

In this paper, the diagnoses of SK and FW are taken as examples. We collect 755 SK images and 545 FW images, which include 43 labeled feature images and 48 labeled non-feature images of SK, and 44 labeled feature images and 41 labeled non-feature images of FW. The incremental experiment is a process of automatic labeling by machine. In this process, our goal is to label the feature images to expand the training data set. To ensure the experimental performance, the number of images cannot be too

Discussion and conclusion

Previously, skin disease research focused on image processing technology to identify the categories of diseases. There are limited studies on the identification and classification of SK and FW. In this paper, confocal laser scanning microscope images with varying degrees of SK and FW are used. However, the machine and the doctor are different, and the machine creates differences in the image's light intensity, shooting depth and angle. This interference makes the image more complex in our

Acknowledgments

This work is supported by Natural Science Foundation of China (61672535, 61502540), Key Laboratory of Information Processing and Intelligent Control of Fujian Innovation Fund (MJUKF201735). The authors declare that they have no conflict of interests.

Kehua Guo is a professor at the School of Information Science and Engineering, Central South University, China. He received the Ph.D. degree in Computer Science and Technology from Nanjing University of Science and Technology. His research interests include artificial intelligence and, ubiquitous computing.

References (30)

  • Y. Zhang et al.

    A survey on emerging computing paradigms for big data

    Chin. J. Electron.

    (2017)
  • B. Farahani et al.

    Towards fog-driven IoT eHealth: promises and challenges of IoT in medicine and healthcare

    Future Gener. Comput. Syst.

    (2018)
  • T. Kooi et al.

    Large scale deep learning for computer aided detection of mammographic lesions

    Med. Image Anal.

    (2017)
  • X. Zhou et al.

    Multi-dimensional attributes and measures for dynamical user profiling in social networking environments

    Multimed. Tools Appl.

    (2015)
  • X. Wang et al.

    A tensor-based big service framework for enhanced living environments

    IEEE Cloud Comput.

    (2016)
  • N. Zhong et al.

    Research Challenges and Perspectives on Wisdom Web of Things (W2T)

    (2016)
  • X. Wang et al.

    A cloud-edge computing framework for cyber-physical-social services

    IEEE Commun. Mag.

    (2017)
  • X. Zhou et al.

    Organic streams: data aggregation and integration based on individual needs

  • A. Esteva et al.

    Dermatologist-level classification of skin cancer with deep neural networks

    Nature

    (2017)
  • Y. Bar et al.

    Chest pathology identification using deep feature selection with non-medical training

    Comput. Methods Biomech. Biomed. Eng.

    (2016)
  • M. Turan, Y. Almalioglu, E. Konukoglu, et al., A Deep Learning Based 6 Degree-of-Freedom Localization Method For...
  • N.J. Dhinagar et al.

    Early diagnosis and predictive monitoring of skin diseases

  • K.P. Kyriakis et al.

    Lifetime prevalence fluctuations of common and plane viral warts

    J. Eur. Acad. Dermatol. Venereol.

    (2007)
  • J.M. Yeatman et al.

    The prevalence of seborrhoeic keratoses in an Australian population: does exposure to sunlight play a part in their frequency?

    Br. J. Dermatol.

    (1997)
  • I. Zaqout

    Diagnosis of skin lesions based on dermoscopic images using image processing techniques

    Int. J. Signal Process. Image Process. Pattern Recognit.

    (2016)
  • Cited by (15)

    • IoT-based telemedicine for disease prevention and health promotion: State-of-the-Art

      2021, Journal of Network and Computer Applications
      Citation Excerpt :

      The latter is accomplished by providing high-resolution retinal images (Jebadurai and Dinesh Peter, 2018). A cognitive system based on a deep convolutional neural network is capable of providing diagnostic advice to medical specialists (Guo et al., 2018). Decentralised services in the domain of healthcare can minimise access to emergency rooms for erroneous therapies or for therapies not adapted to changes in patients’ medical records (Patti et al., 2018).

    • Deep tensor fusion network for multimodal ground-based cloud classification in weather station networks

      2020, Ad Hoc Networks
      Citation Excerpt :

      However, hand-crafted cloud features are difficult to generalize to large-scale databases. In recent years, convolutional neural networks (CNNs) have achieved excellent performance in wireless sensor networks [4,5], computer vision [6–8], remote sensing [9,10] and so on. The advantage of CNN-based methods lies in automatically learning features according to the distribution of large-scale database.

    • iFusion: Towards efficient intelligence fusion for deep learning from real-time and heterogeneous data

      2019, Information Fusion
      Citation Excerpt :

      Data is stored in different locations, information asymmetry between different departments and enterprises, and lack of sharing channels, which lead to inefficient use of a large number of data resources. In addition, some data are related to privacy issues [14], such as medical data containing the patient's private information [15], and these data cannot be shared. Therefore, to ensure that deep learning enables improved handling of real-time data and heterogeneous data, we propose an intelligence fusion framework called iFusion to address the above problems.

    • Towards efficient medical lesion image super-resolution based on deep residual networks

      2019, Signal Processing: Image Communication
      Citation Excerpt :

      A CNN can greatly reduce the number of parameters in the neural network and simultaneously reduce the complexity of the neural network model [7]. Based on a CNN, the classification of dermatological images generated by confocal laser scanning microscopy is almost as accurate as that performed by individual dermatologists [8]. In the field of super-resolution (SR), CNNs have significantly improved the peak signal-to-noise ratio (PSNR) and the structural similarity (SSIM) index [9].

    • IOT-Based Telemedicine Network Design Implementation of a Smart Health Monitoring System in COVID-19

      2023, Practical Artificial Intelligence for Internet of Medical Things: Emerging Trends, Issues, and Challenges
    View all citing articles on Scopus

    Kehua Guo is a professor at the School of Information Science and Engineering, Central South University, China. He received the Ph.D. degree in Computer Science and Technology from Nanjing University of Science and Technology. His research interests include artificial intelligence and, ubiquitous computing.

    Ting Li received the B.SC. degree in Computer Science and Technology from Xiangtan University in 2016. She is currently a graduate student at the School of Information Science and Engineering, Central South University, China. Her research interests include ubiquitous computing and big data.

    Runhe Huang received Ph.D. in Computer Science and Mathematics from the University of the West of England. She is a professor in the Faculty of Computer and Information Sciences at Hosei University, Japan. Her research fields include computational intelligence, and ubiquitous intelligence computing.

    Jian Kang received the Ph.D. degree in Medicine from the Nanjing Medical University, China. He is a doctor in the Department of Dermatology of the Third Xiangya Hospital of Central South University, China. His research fields include information and network management and diagnosis of dermatosis.

    Tao Chi received the Ph.D. degree in Tongji University, China. He is a professor in the Key Laboratory of Fisheries Information, Ministry of Agriculture, Shanghai Ocean University. His research fields include Internet of Things and ubiquitous computing.

    View full text