Deep learning in medical imaging: A brief review

Abstract

Researchers have used deep learning methods for a human level or better disease identification and detection. This paper reports, in brief, the recent work in deep learning identification of diseases occurring at three unique parts of the human body: the skin, the thorax, and the eye. While earlier reviews reported on the theory, applications, and challenges of such research, what distinguishes this work from the others is the reporting and comprehensive analysis of the key results. In doing so, the paper not only summarizes the major conclusions drawn from them but also emphasizes their weaknesses. The hope is to help the researchers see the big picture in deep learning classification of the diseases of the skin, the thorax and the eye, and guide them to find the right future research direction. INTRODUCTIONPhysicians use different imaging modality techniques, such as MRI, CT, ultrasound, and X-ray to help them detect, diagnose, and cure serious diseases. With help from radiologists, they can interpret a vast number of diseases occurring at distinct parts of a human body, ranging from brain to lungs to skin. Still, the tremendous diversity in disease pathology can make human errors possible.In search of computer methods to aid in disease detection and recognition, researchers have experimented on convolutional neural networks (CNN) since the mid-90s. 1 Used for hand-written digit recognition at first, 2 CNN's popularity increased with the discovery of deep CNNs in 2012. [3][4][5][6][7][8] Since then, a growing number of researchers are using CNNs for medical image analysis (such as References 9-17). Some even show that CNNs can give comparable performance to medical doctors. 18 We may use the Internet of Things (IoT) to connect medical devices in hospitals for the collection of big data. Each hospital can join a central unit this way for the automated detection of diseases on images using CNN. IoT is a crucial tool for reaching data in an instant, and artificial intelligence is essential for modeling for disease detection on images. Together they may form what we call The Internet of Medical Things (IoMT). 19 CNN's distinct architectures and vast application areas (such as this and References 20-22) have likewise led the research groups to publish a few review papers on using deep CNNs in medical imaging (eg,. While these papers focus on the theory, applications, and challenges of deep CNNs, they do not detail the results discovered by the researchers.This survey paper, yet, aims to fill that gap by reporting the recent advances of deep learning use in medical imaging and analyzing the key results. In doing that, it hopes to offer the reader an opportunity to see the big picture, including any strengths and weaknesses that such research might have. The intention is not to cover the entire application areas in medical imaging but focus on only a few parts of the human body and be thorough. Hence, this survey paper includes

Artificial intelligence in positioning between mandibular third molar and inferior alveolar nerve on panoramic radiography

“…ResNet-50, mainly used for medical image classification 22 , is a substantially deeper and easier model to train compared to simple models such as VGGnet, and the core structure is a residual block 23 . Residual learning does not allow for error accumulation on the convolution layers but enables a better representation of the content in the convolution layers.…”

Artificial Intelligence in Positioning Between Tooth and Nerve on Panoramic Radiography

“…ResNet-50, mainly used for medical image classi cation 19 , is a substantially deeper and easier model to train compared to simple models such as VGGnet, and the core structure is a residual block 20 . Residual learning does not allow for error accumulation on the convolution layers but enables a better representation of the content in the convolution layers.…”

Cancer-Cell Deep-Learning Classification by Integrating Quantitative-Phase Spatial and Temporal Fluctuations

“…ResNet-50 solves this problem by identifying shortcut connections, skipping certain layers while providing great generalization performance with a relatively small number of parameters. Indeed, ResNet-50 has been successfully used for many medical image classification tasks [32,33]. We therefore selected it for our case as well.…”