User:Yasmin2023/Sample page

= Study of deep learning techniques for medical image analysis: A review =

Ayush Singhal (a), Manu Phogat (a), Deepak Kumar (a) , Ajay Kumar (b), Mamta Dahiya (a), Virendra Kumar Shrivastava (c)
 a. Department of Computer Science Engineering, Faculty of Engineering and Technology, Shree Guru Gobind Singh Tricentenary University, Gurugram 122505, Haryana, India 

 b. Department of Mechanical Engineering, Faculty of Engineering & Technology, Shree Guru Gobind Singh Tricentenary University, Gurugram 122505, Haryana, India 

 c. Department of Big Data Analytics, Adani Institute of Digital Technology Management, Gandhinagar, Gujarat 382423, India 

Abstract
Presently, there is a significant surge of research activity in the realm of medical image analysis, with a primary emphasis on the utilization of deep convolutional networks. Deep learning employs a diverse array of models designed to discern and extract vital information from the images fed into these sophisticated neural networks. This transformative technology has found widespread adoption in the field of medicine, where it serves as a valuable tool for disease detection and diagnosis, and subsequently, for classifying diseases into specific categories. Notably, one of the most widely adopted models for medical image analysis is the Convolutional Neural Network (CNN). In essence, the core focus of this review paper revolves around the application of deep learning, particularly the use of deep neural networks, to the task of disease detection. This process involves the retrieval and extraction of critical information from the medical images provided as input to the network. The paper goes beyond this technical aspect and also provides insights into the practical, real-world applications of deep learning within the medical field. It delves into how this technology is harnessed in clinical settings to improve disease diagnosis and patient care.

Furthermore, the review paper doesn't shy away from addressing the limitations and challenges inherent in the use of deep learning for image analysis in the medical domain. It highlights the complexities and potential drawbacks associated with this approach, shedding light on areas where further research and development are needed to enhance its effectiveness. Importantly, it's worth noting that this paper is copyrighted by Elsevier Ltd. (© 2022), with all rights reserved. It has undergone a rigorous selection and peer-review process overseen by the scientific committee of the International Conference on Materials, Machines, and Information Technology-202. This underscores the paper's credibility and quality within the academic and research community.

Introduction
Deep learning has made significant progress across various fields, including medicine. It is now essential for drug innovation, clinical decisions, and medical imaging. With the transition to digital health records, the role of medical imaging in maintaining patient data is crucial. Traditionally, human radiologists interpreted these images, but training them is costly and time-consuming. Automated deep learning solutions are increasingly necessary for accurate and efficient image analysis in the medical sector.

Deep learning excels with high-dimensional data and has advanced image and speech recognition, drug prediction, genetic analysis, and disease forecasting. In medicine, it is applied to detect various disorders, such as diabetic retinopathy, tumors, and Interstitial Lung Disease, often using Convolutional Neural Networks (CNNs). Specialized tools enhance the efficiency of healthcare professionals, assisting in tasks like chest radiograph orientation detection and identifying cellular components in pathology slides. Medical imaging is vital for disease detection and diagnosis, particularly when frequent screenings and immediate expert availability are not possible.

This review paper covers:


 * 1) Types of Medical Images: Discusses various digital medical imaging methods.
 * 2) Medical Image Analysis History: Provides a historical context for image analysis.
 * 3) Deep Learning Methods: Explores the role of CNNs and other models in medical image analysis.
 * 4) Literature Survey and Datasets: Reviews existing studies and datasets.
 * 5) Limitations in Deep Learning and Future Prospects: Addresses challenges and potential solutions.
 * 6) Conclusion: Summarizes the significance of medical image analysis in healthcare.

Deep learning enhances the accuracy and efficiency of medical imaging, transforming disease diagnosis and healthcare accessibility.

Applications
The Convolutional Neural Network (CNN) model is employed in various medical applications to classify diseases, and these applications are associated with specific datasets. Here's a summary:


 * 1) Lung Texture Classification using CT Scans:
 * 2) * Modality: CT Scans
 * 3) * Dataset: 73 CT scans with 5 different classes [39]
 * 4) Lung Pattern Classification using High-resolution CT Scans:
 * 5) * Modality: CT Scans
 * 6) * Dataset: 109 high-resolution CT scans [40]
 * 7) Breast Cancer Classification using Histology Images:
 * 8) * Modality: Histology images
 * 9) * Dataset: BreakHis [41]
 * 10) Breast Cancer Diagnosis using Mammographic Images with ROIs (Region of Interests):
 * 11) * Modality: Mammographic Images
 * 12) * Dataset: 3158 ROIs categorized into 2 classes [42]
 * 13) Brain Tumor Detection (Cancer Detection) using MRI Images:
 * 14) * Modality: MRI images
 * 15) * Dataset: BraTS [43]
 * 16) Diabetic Retinopathy Classification using Retina Images (Kaggle Dataset):
 * 17) * Modality: Retina images
 * 18) * Dataset: 80,000 images classified into 5 classes [44]
 * 19) Skin Tissue Classification using Dermo.S. Images:
 * 20) * Modality: Dermo.S. images
 * 21) * Dataset: ISIC [45]
 * 22) Liver Tumor Segmentation using CT Slices:
 * 23) * Modality: CT Slices
 * 24) * Dataset: CT Slices [46]
 * 25) Dermoscopy Patterns Classification using Dermo.S. Images:
 * 26) * Modality: Dermo.S. images
 * 27) * Dataset: ISIC [47]
 * 28) Nuclei Classification using Histology Images of Colorectal Adenocarcinomas:
 * 29) * Modality: Histology images
 * 30) * Dataset: 100 histology images of colorectal adenocarcinomas [48]

Types of medical imaging
Medical imaging refers to the methodology and procedure used to create internal representations of the human body for clinical examination and the identification of anatomical components. This field encompasses radiography, which employs various imaging technologies like X-Rays, PET scans, MRIs, SPECT scans, and dermoscopy images to diagnose patients. These imaging modalities can either focus on individual body parts sequentially or simultaneously examine multiple organs. In the medical sector, the utilization of these imaging modalities is progressively growing. Moreover, each of these modalities generates varying output data. For example, MRI scans produce data files that can be quite large, often reaching hundreds of megabytes, while histology slides yield smaller datasets.

The CT scan technique is employed for the diagnosis of internal organs, bones, and blood vessels. It operates similarly to X-Rays but captures images from a 360-degree perspective around the patient. In CT scan images, high-density structures like bones appear white, while low-density structures appear black. MRI, on the other hand, is akin to CT scans but provides more intricate and detailed information. It is particularly useful for diagnosing and visualizing organs, including the brain and other bodily tissues. Positron Emission Tomography (PET) serves the unique purpose of providing insight into organ function rather than just structural information. This method generates 3D images of the body's interior and is particularly valuable for diagnosing diseases like cancer.

Medical image analysis history
In the 1970s, a groundbreaking Artificial Intelligence prototype marked the beginning of rule-based, expert systems, offering a significant development in the field. Notably, in the realm of medical science, the MYCIN expert system, developed by Shortliffe, played a pivotal role in recommending various antimicrobial treatment options for patients. As the field of AI evolved, it transitioned from heuristic-based methods to manually crafted feature extraction strategies and eventually to supervised learning techniques. While unsupervised machine learning methods were explored, there was a notable shift towards supervised algorithms between 2015 and 2016, with a particular focus on models like convolutional neural networks (CNNs) by 2017.The foundational principles of artificial neurons can be traced back to the work of McCulloch and Pitts in 1943, which laid the groundwork for the development of the perceptron in 1958. Artificial neural networks, consisting of interconnected data processing units called neurons, form the basis of deep neural networks. Deep neural networks, similar to artificial neural networks (ANN), feature multiple layers. These deep layers have the capacity to automatically learn low-level features, such as curves, edges, and vertices. At higher or more abstract levels, these networks combine features extracted from lower-level layers to identify image classes or object shapes. This hierarchical process closely resembles how the human brain's cortical region processes visual data and recognizes objects.

In 1982, Fukushima introduced the concept of the Neocognitron, an early precursor to CNNs. However, it was Lecun et al. who generalized and popularized CNNs. They applied error backpropagation in an experiment to recognize handwritten digits. CNNs gained widespread adoption when Krizhevsky et al. introduced key concepts in 2012, including non-linear activation functions, ReLU functions, dropout algorithms, and more. Notably, in 2012, they achieved first place in the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) competition with a CNN that achieved a remarkable 15% error rate, surpassing the second-place model by a substantial margin with a 26% error rate. Following this success, CNNs have continued to dominate ILSVRC competitions since 2015, consistently outperforming human performance in image recognition. Consequently, there has been a significant increase in literature and applications related to CNNs, underscoring their growing importance in the field of disease diagnosis.

1. Convolutional neural network
CNNs are like specialized tools for understanding pictures, but they can also help with sorting and understanding data. These networks have something called "convolutional layers" that make them special. These layers, along with other parts, build a complicated network. Each layer in this network can recognize important things in a picture. CNNs use a trick called "filtering" to figure out these important things. It's like looking at different parts of a picture through a special lens to see how well they match with the filter. This whole process is called "convolution." Each filter makes a map of what it found. When we move to higher layers, these maps help us learn even more details.

Image feature extraction using convolution neural network

 * 1) Convolution Layer
 * 2) * The convolution layer is a part of CNN used to find patterns in images. Imagine an image made up of tiny dots called pixels. A CNN learns from these images to understand patterns and recognize objects. It looks for special patterns using something called a "feature detector" or "filter."
 * 3) * A feature detector is like a small grid with numbers. The computer sees the image as a grid of numbers, where each number represents the brightness of a pixel. When you apply the feature detector to the image, it does some math with the numbers. This math helps the CNN find important features in the image, like edges or textures. As the feature detector moves across the image, it creates a new map with this information.
 * 4) * The CNN then uses this new map in the next step. It adds up all the numbers in the map, adjusts with a bias value, and applies a function to get a final result.
 * 5) ReLU Layer
 * 6) * The ReLU layer is another part of the CNN that helps with image understanding. Images are complex, and objects in them can have different shapes and colors. When the CNN looks at images, it can create linear results. To make it understand non-linear parts of images, we use the ReLU layer.
 * 7) * ReLU stands for "Rectified Linear Unit." It's a type of function that is linear for positive values and zero for negative values. This function helps the CNN capture non-linear features in the images, like edges and shapes. It's a good choice because it speeds up the learning process and improves image classification.
 * 8) Pooling Layer
 * 9) * After the convolution, we have the pooling layer. This layer reduces the size of the image and simplifies it. Different types of pooling can be used, but the most common is the max-pooling layer.
 * 10) * In max-pooling, we take the largest number from a small part of the image and ignore the others. This is done in blocks, and we collect all the max numbers to create a smaller image. It helps to reduce the amount of data and focus on the most important information.
 * 11) Fully Connected Layer
 * 12) * The fully connected layer is the last step in the CNN. It helps make decisions about the image. It combines the information from previous layers to make sense of what's in the picture.
 * 13) * This layer links every piece of information to each other, so it can understand the image as a whole. In cases like image classification or object detection, it's this layer that tells us what's in the picture.
 * 14) Dropout Regularization
 * 15) * Overfitting is a problem in deep neural networks, which happens when a network becomes too specialized and fails to work well with new examples. Dropout is a technique to prevent this.
 * 16) * In dropout, we randomly turn off some connections in the network during training. This means that some parts of the network learn less from the data. It helps to prevent overfitting and make the network more versatile. It's like keeping a balance between old knowledge and new knowledge to build a better model.

2. Recurrent neural network
Recurrent Neural Networks (RNNs) are a type of neural network that's great at recognizing patterns in data that comes in sequences. This data could be in the form of text, voice, images, music, genetic sequences, or even events in a medical setting.

In a regular neural network, all the inputs and outputs are independent of each other. But when you need to predict the next word in a sentence, you have to consider the words that came before it. RNNs have a hidden layer that helps them remember these sequences, making them really good at predicting the next word in a sentence.

Training RNNs can be tricky because of a problem called "gradient vanishing" and "exploding." To address this, a more advanced version called Long Short-Term Memory (LSTM) was introduced. LSTM can remember sequences over longer periods.

In the field of radiology, where sequential data like ultrasound videos are common, RNNs are widely used. Radiologists use them, for example, for transcribing medical images into text reports. There are also cases where RNNs are combined with other models like Convolutional Neural Networks (CNNs) to analyze complex image data, like electron microscope images for identifying fungal and neuronal structures.

3. Autoencoders
Autoencoders are a type of unsupervised learning method used for representation learning. What they do is pretty clever. They learn how to take some data and make it smaller and simpler (that's the encoding part). Then, they figure out how to take that simplified data and turn it back into the original data (that's the decoding part). This helps in getting rid of any unnecessary or noisy bits in the data. Autoencoders have four important parts:


 * 1) Encoder: This part learns how to make the data smaller and simpler.
 * 2) Bottleneck: It's like a tiny storage space where the really compressed data is kept.
 * 3) Decoder: The decoder's job is to take that tiny data and turn it back into something that's very close to the original.
 * 4) Reconstruction loss: This checks how well the decoder is doing its job. It looks at how close the decoded data is to the original data.

Literature review and datasets
Deep learning is a powerful tool in medicine, assisting with drug development and patient care decisions. It's particularly valuable for analyzing intricate medical images.

Examples of deep learning in medicine:


 * 1) Lo's team used deep learning in 1995 to identify lung nodules in chest X-rays using a CNN with two hidden layers.
 * 2) Rajpukar's team employed a deep CNN with 121 layers to classify 14 chest infections from a dataset of 112,000 chest X-ray images.
 * 3) Shin's group used stacked autoencoders to locate kidney, liver, and heart in abdominal MRI images.
 * 4) Akkus and team conducted a comprehensive study on MRI brain segmentation, exploring various CNN models and metrics.
 * 5) Small filters in deep learning networks simplified MR image analysis and prevented overfitting.
 * 6) 3D patches and an 11-layer network with special techniques were used to identify brain lesions.
 * 7) A specialized deep learning model classified lung images into five groups, addressing data limitations and overfitting with dropout techniques.
 * 8) Deep learning was applied to recognize body parts in a two-step process: CNN feature extraction and network adjustment based on these features.

These experiments demonstrate deep learning's significant role in medical image analysis and data processing.

Limitations in deep learning and future Application
Deep learning is fantastic for improving performance, but it has some limitations, especially in clinical applications. These models need a lot of data to train, and sometimes you need labeled data, which is tough to do manually. However, as deep learning technology and digital storage for medical images get better, this problem is getting easier to manage. There are other limitations, like noise in medical images. But you can fix that with some pre-processing steps and techniques like transfer learning. Transfer learning lets deep models work well with smaller datasets. In medical imaging, they also use GANs when there's not much image data available. And there's a new deep learning model called Capsule Network (CapsNet) that's trying to solve some of the problems of CNN models.

Conclusion
In conclusion, deep learning methods have shown remarkable performance across various areas of medical image analysis, such as disease detection and classification. These advancements are set to enhance the accuracy and efficiency of computer-aided diagnoses. We can anticipate further research and adaptations of this technology to be applied in other medical fields where it hasn't been used before. The overall success of deep learning in medical image analysis suggests significant benefits for both the healthcare industry and patients. The authors also confirm that they have no competing financial interests or personal relationships that could have influenced the work presented in this paper.