Deep learning is showing a growing trend in data analysis and is known as one of the 10 breakthrough technologies in 2013 [1]. It is an improvement of the neural network, including more layers of computation, enabling higher levels of abstraction and prediction in the data [2]. So far, it is becoming the leading machine learning tool in the field of general imaging and computer vision.
In particular, Convolutional Neural Networks (CNN) have proven to be an advantageous tool for many computer vision tasks. Deep CNN automatically learns intermediate and advanced abstractions derived from raw data (eg, images). Recent results indicate that the generic descriptor extracted from CNN is very effective in object recognition and localization of natural images. Medical image analysis groups around the world are rapidly entering the field and applying CNN and other deep learning methods to a wide range of applications. Many good results are emerging.
In the field of medical imaging, accurate diagnosis or assessment of disease depends on image acquisition and image interpretation. In recent years, with the development of technology, devices can collect data at a faster rate and more powerful resolution, which greatly improves the quality of image acquisition. However, improvements in image interpretation by computer technology are just beginning. Currently, most medical image interpretations are performed by doctors. However, image interpretation by humans is often one-sided because of its subjectivity, large changes in different interpreters, and fatigue. Many diagnostic tasks require an initial search process to detect anomalies and quantify changes in measurements and time. Computerized tools, especially image analysis and machine learning, play a key role in improving diagnosis. They support expert workflows by helping identify areas that require treatment. Among these tools, deep learning is quickly confirmed as a basis for superiority and accuracy. It has also opened up new areas of data analysis and is evolving at an unprecedented rate.
A. Historical network
The basic ideas behind neural networks and deep learning have existed for decades [3]. They usually have only a few layers. The emergence of backpropagation algorithms has led to a significant increase in the performance of neural networks. However, performance is still not enough. Other classifiers have evolved, including decision trees, boosTIng, and support vector machines. Each of them has been applied to medical image analysis, especially for detecting anomalies, and they have also been applied in other related fields such as segmentaTIon. Despite this development, relatively high false positive rates are still common.
As early as 1996 in the work of Sahiner et al., CNN (Convolutional Neural Network) was applied to medical image processing [4]. In this work, ROIs (Region of Interests) containing biopsy-confirmed masses or normal tissues were extracted from mammograms. The CNN contains an input layer, two hidden layers, and an output layer, as well as backpropagation used. In this pre-GPU era, training time was described as "calculatively intensive" but did not give specific time. In 1993, CNN was used for lung nodule detection [5]. In 1995, CNN was used to detect microcalcifications in mammograms [6].
A typical CNN for image processing consists of a series of layers of convolution filters interspersed with a series of data compression or pooling layers. A convolution filter (convoluTIon filter) processes a small block of the input image. Similar to low-level pixel processing of the human brain, convolution filters can detect highly correlated image features, such as lines or circles that can represent sharp edges (for example, for organ detection) or circles (such as objects for circles, Like colon polyps, then high-order features such as local or global shapes and textures. The output of the CNN is typically a label of one or more probabilities or categories corresponding to the image. The convolution filter can learn directly from the trained data. This is exactly what people need because it reduces the need for time-tagged features that take time. If a convolution filter is not used, then in the pre-processed image phase, filters designed for a particular application and some features that need to be computed are inseparable from these artifacts.
CNN is a highly parallelized algorithm. A large part of the utility of using CNN is derived from the huge speed increase (about 40 times) that is facilitated by the image processing unit (GPU) compared to separate CPU processing. An early paper describing the value of GPUs for training CNN and other machine learning techniques was published in 2006 [8]. In medical image processing, GPUs are first introduced for segmentation, reconstruction, and registration, and then machine learning [9], [10]. Interestingly, although Eklund et al. [10] talked extensively about convolution in their 2013 paper, convolutional neural networks and deep learning were not mentioned at all. This highlights how rapidly the major reforms in deep learning have rapidly adjusted medical image processing research.
B. Today's network
Deep neural networks have recently gained considerable commercial interest due to the development of new variants of CNN and the emergence of efficient parallel solvers for modern GPU optimization. The power of CNN benefits from its deep architecture, which allows it to extract a range of distinguishing features at different levels of abstraction. Training a deep convolutional neural network from scratch is a huge challenge. First, CNN requires a large amount of tagged data, which is difficult to achieve in the medical field. This is because it is very expensive to ask an expert to mark it, and samples of diseases (such as lesions) are very rare. Second, training deep CNN requires a lot of computational and memory resources. Without them, the training process can be very time consuming. Third, training a deep CNN often complicates due to overfitting and convergence problems, often requiring repeated adjustments to the learning parameters or architecture of the network to ensure that all layers learn at a fairly speed. In view of the above difficulties, some new learning programs called "transfer learning" and "fine-tuning" have been proposed to provide solutions and are accepted by more and more people. These will be discussed further in Section II-C.
C. Network in the medical field
The domain deep learning method is most effective when applied to large training sets, but in the medical field, it is not always possible to obtain large data sets. Therefore, we face a series of major challenges, including: (a) Can deep neural networks be used effectively in medical tasks? (b) Is it relevant to transfer learning from general imagery to the medical field? (c) Can we rely solely on the characteristics of learning, or can we combine them with hand-crafted functions to accomplish the task? This IEEE Imaging (IEEE-TMI) special problem for deep learning of medical imaging focuses on the advancement of this new era of machine learning and its role in the field of medical image processing. This question describes the recent achievements of CNN and other deep learning applications in medical tasks. It contains 18 articles selected from 50 papers from various investigators around the world. This is a very high number for the IEEE special issue, and this is the length of time from the publication of the paper to the submission deadline. It used to be realized in a short period of time. The paper focuses on a large number of traditional tasks from detection to categorization (eg, lesion detection, image segmentation, shape modeling, image registration), as well as open and novel applications. It also includes some work focused on network exploration, and gives a perspective on how different tasks, parameters, and training sets should be chosen.
Overview of journal articles and topics
A. Disease detection
CAD/Computer-aided detecTIon is a well-established field of medical image analysis that is ideal for deep learning. In the standard method of CAD [11], lesions are detected by supervised methods or classical image processing techniques such as filtering and mathematical morphology. Candidate lesions are usually segmented and are usually described by a large number of manually designed features. The classifier maps the feature vector to the probability that the corresponding candidate site is the actual lesion. A straightforward approach to using deep learning rather than manually designed features is to train CNNs that operate on a set of image image data centered on candidate lesions. This article is used in several articles in this issue. In order to obtain candidates for lung nodules from 3D chest CT scans and to extract 2D patches in 9 different directions centered on these candidate sites, Setio et al. [12] combined three previously developed candidate probes. Each candidate is classified by combining different CNNs. The report shows that this approach has made minor progress compared to the results of previously published classic CAD for the same task.
Roth et al. [13] used CNN to improve three existing CAD systems for detecting the presence of colonic polyps in colonoscopy CT and for detecting sclerotic spinal metastases and enlarged lymph nodes using body CT. They also used previously developed candidate detectors and 3 orthogonal 2D patches, as well as up to 100 random rotation views. A randomly rotated "2.5D" view is a method of decomposing an image from the original 3D data. Additional accuracy gains are then obtained by integrating the predictions of CNN on the 2.5D view. For all 3 CAD systems using CNN, the sensitivity of lesion detection improved by 13 – 34%, indicating that the method is universally adjustable. If a non-deep learning classifier (such as a support vector machine family) is used, it is almost impossible to achieve this level of improvement.
Dou et al. [14] detected cerebral microbleeds from magnetic susceptibility-weighted magnetic resonance imaging scans. They used a 3D CNN and replaced the original candidate detection phase with CNN, proposing a two-stage approach. The report suggests that the results of their 3D CNN are re-implemented, trained, and tested on the same dataset, as compared to other classical methods and 2D CNN methods in the existing literature.
Sirinukunwattana et al. [15] detected and classified nuclei in histopathological images. They use a CNN, using a small block as input, not just predicting whether the central pixel of the small block is a nucleus, they model the output, and each cell core produces a peak and is flatter elsewhere. This spatially constrained CNN combined with the fusion of overlapping patches in the test phase yielded better results than previously proposed techniques based on CNN and classical feature based methods.
Anthimopoulos et al. [16] focused on the use of 2D chest CT scan images to detect patterns of interstitial lung disease. They are one of three groups that study this problem (the other two are Shin et al. [17] and van Tulder et al. [18]) using a public data set from [19]. They train the CNN so that it can distinguish which of the 7 classes the small block of 32 × 32 pixels belongs to. The report shows that their results have achieved higher precision than the previous three methods using manual design features.
In several other articles on such issues, lesion detection is also a topic of interest, but the focus of these articles is on broader or focused on specific methodological issues. These papers will be briefly discussed below.
B. Segmentation and shape modeling
For a large data set consisting of 2891 cardiac ultrasound examinations, Ghesu et al. combined deep learning and marginal space learning for object detection and segmentation. The combination of "effective exploration of large parameter space" and a method of enhancing sparsity in deep networks enhances computational efficiency, and the method reduces the average segmentation error by 13.5% compared to another reference method published by the same group. .
Three groups of researchers focused on brain structure segmentation or brain lesions. Magnetic resonance imaging (MRI) problems in the segmentation of multiple sclerosis brain lesions are addressed by Brosch et al. They developed a 3D deep convolutional coding network that combines the concatenated convolution and deconvolution processes. The convolution process learns a higher level of functionality, while the deconvolution process predicts voxel level segmentation. They applied this network to two public data sets and a clinical trial data set and compared their methods to five common methods. According to the report, the performance of the method is “comparable to the most advanced methods availableâ€.
Pereira et al. studied brain tumor segmentation in magnetic resonance imaging. [ twenty two ]. They use a small kernel, deeper architecture, grayscale normalization, and data enhancement. Different convolutional neural network architectures are used for low-level and high-level tumors. The method separates the enhanced portion and the core portion of the tumor, respectively. They ranked first in the 2013 Dataset Public Challenge and ranked second in the 2015 Live Challenge.
For brain segmentation, a study by Moeskops et al. showed that convolutional neural networks performed well in data sets covering five different age groups from premature to older. A multi-scale approach is used to achieve its reliability. The method achieved good results in 8 tissue categories, with an average of Dice similarity coefficients for the 5 data sets ranging from 0.82 to 0.91.
C. Network inquiry
1) Data Dimensions - 2D vs 3D: Most of the data we have seen is based on 2D analysis. The two-dimensional transition to three-dimensional is often questioned—whether it will be a key to a significant performance boost. There are some changes in the data enhancement process, including 2.5 dimensions. For example, in the study by Roth et al., axial, coronal, and sagittal images centered on candidate colon polyps or voxels in lymph nodes and were imported into the cuda-convnet convolutional neural network, which included A red, green, and blue channel that is commonly used to represent a natural light image. The three-dimensional convolutional neural network was explicitly used by Brosch and Dou et al.
2) Learning Methodology - Unsupervised vs. Supervised: When we look at the online literature, it is obvious that most studies focus on supervised convolutional neural networks in order to achieve classification. This type of network is very important for many applications, including detection, segmentation, and tagging. However, there are still some studies focusing on unsupervised schemes, which have proven to be useful in image coding, efficient image characterization schemes, and as a pre-processing step in an in-depth surveillance approach. Unsupervised characterization learning methods such as Restricted Boltzmann Machines (RBM) may go beyond standard filter banks because they learn characterization directly from training data. RBM is trained by a generative learning objective; this allows the network to learn characterizations from unmarked data, but does not necessarily produce the best classification features. Van Tulder et al. conducted a survey that combines the advantages of generative and discriminative learning objectives in volume-based RBMs. This study shows that the combination of learning tasks is superior to pure discriminative or generative learning.
3) Training data considerations: The convolutional neural network implements the learning of data-driven, highly characterized, hierarchically hierarchical image features. In many applications (see journals), these features have proven to be a very powerful and reliable representation. To provide such a rich representation and successful classification, sufficient training data is required. The amount of data required is a key issue to be explored. Related questions include the following: How do we most effectively use the training data we have? What can we do without data? And finally, is there an alternative way to get data and make medical annotations?
Some of these issues are addressed by some of the papers in the journal. Van Grinsven et al. attempted to improve and accelerate convolutional neural network training for medical image analysis tasks by dynamically selecting negative samples that were misclassified during training. The convolutional neural network training process is a continuous process that requires multiple iterations (or multiple periods) to optimize network parameters. At each time period, a subset of samples is randomly selected from the training data and presented to the network to update its parameters by backpropagating and minimizing the cost function. The classification task in the medical field is often a normal/pathological discrimination task. In this case, the normal type category is especially over-characterized; in addition, most normal-type training samples are highly correlated due to the repetitive pattern of normal tissue in each image. Only a small percentage of them will contain useful information. Equivalent processing of these data during the learning process can result in wasting many training iterations on useless normal samples, making the training process of the convolutional neural network take unnecessary time. One method of identifying useful normal samples (as shown in the study) improves the efficiency of the convolutional neural network learning process and reduces training time.
4) Migration learning and fine-tuning: Getting data in the field of medical imaging as fully annotated as ImageNet remains a challenge. There are several ways to help when there is not enough data available: 1) Migration learning: by a natural image dataset or by a convolutional neural network model (supervised) pre-trained in a different medical field Used for a new medical task at hand. One solution is that a pre-trained convolutional neural network is applied to an input image and its output is extracted from the network layer. The extracted output is treated as a feature and used to train a separate pattern classifier. For example, in a study by Bar et al., a pre-trained convolutional neural network was used as a feature generator to identify chest pathology. In the study by Ginneken et al., features based on convolutional neural networks are integrated with manually added features to achieve performance improvements in a nodule detection system. 2) Fine-tuning: There is indeed a medium-sized data set available for the task at hand. One reference scheme is to use a pre-trained convolutional neural network as several (or all) network layers after further supervised training is completed. Initialize to use, use new data in the task at hand.
Migration learning and fine-tuning are key components of the use of deep convolutional neural networks in medical imaging applications. The discussion of these issues is the work of Shin and Tajbakhsh et al. The experimental results in the study consistently show that the best results can be achieved using a pre-trained neural network with fine-tuning, whether it is a specific application area (Tajbakhsh et al.) or all network architectures (Shin et al.). Further analysis by Tajbakhsh et al. shows that depth fine-tuning is superior to shallow fine-tuning in performance improvement, and the reduction in the size of the training set also increases the importance of using fine-tuning. In the Shin et al. study, the GoogleNet architecture implements the most advanced mediastinal lymph node detection compared to other shallower depth architectures.
5) Ground Truth - from experts and non-experts: the lack of publicly labeled data, and the difficulty of collecting this data in each medical task, plus cost and time overhead, these are medical areas A prohibitive factor in the middle. Although crowdsourcing implements annotations for large databases of real-world images, its use for biomedical purposes requires a deeper understanding and requires a more precise definition of the actual annotation tasks (Nguyen and McKenna et al. ). The fact that expert tasks are outsourced to non-expert users can lead to cluttered comments that cause disagreements between users. Many problems arise in the combination of medical experts and non-professionals, such as how to combine information sources and how to evaluate and mix input weights by their proven performance in performance and other aspects. These problems are solved by Albarqouni et al. They proposed a network that combines an aggregation layer, which is integrated into a convolutional neural network to make learning input from mass annotations part of the network learning process. The results show valuable insights into the function of deep convolutional neural network learning. The most striking fact about crowdsourcing research in the medical field is the conclusion that a group of non-professional, inexperienced users can actually do as well as medical experts. Nguyen and McKenna et al. also observed this in the study of radiographic images.
D. Innovative applications and novel application cases
Kallenberg [32]'s work is based on breast X-ray images as an input data source, using unsupervised feature learning to score the disease risk of the breast. They show a way to learn hierarchical features from unmarked data, which are then directly imported into a simple classifier. In this classifier, two different types of operations will be performed: 1) image segmentation of breast density, and 2) scoring of mammogram X-slice texture. The classifier performs very well in both areas. To control the capacity of the training model, a sparse regularization optimization is used to control sparse time and range. In the unsupervised learning process, the convolutional layer can be regarded as an automatic encoder autoencoders; in the supervised learning part, the (pre-trained) weights and bias values ​​are further fine-tuned using the softmax regression function.
Yan [33] et al. designed a multi-stage deep learning framework to deal with image classification problems and apply them to human local feature recognition. In the pre-training phase, the convolutional neural network is trained by multi-instance learning to obtain the most discriminative local tiles and local tiles of invalid messages in the current training data slice. In the intensive phase, the pre-trained convolutional neural network will further train the image classifier through the corresponding partial image, thereby enhancing his classification ability. The highlight of this multi-instance deep learning method is that the identification of local images of discriminative partial images and invalid messages can be done automatically. Therefore, no prior manual labeling work is required.
The use of regression networks in medical images is not very common. Miao et al. proposed a regression network based on convolutional neural networks to achieve real-time 2D/3D registration. They proposed three algorithms to simplify the potential mapping object regression and added a strong nonlinear model to the CNN regression model. From the output of this network, the deep learning algorithm is more accurate and robust than the previous optimal algorithm, which greatly improves the gray-based 2D/3D registration process.
We are still exploring the areas in which neural networks can be applied and in which areas their applications and task dimensions will have a lasting impact. In a groundbreaking study, Golkov [35] presented an originalization argument that used deep learning to simplify the processing of diffuse MRI (nuclear magnetic resonance) image data, which was only one step after optimization. Their research shows that this improvement reduces the scan time for scalar measurement data from an advanced model by a factor of 12 and eliminates the need to use diffusion models to identify anomalies. It is worthwhile to reveal the relationship between diffusion-weighted signals and microstructure characteristics. Golkov [35] stated that the use of deep neural networks may reveal such a relationship: Diffusion-weighted imaging (DWIs) can be directly treated as input data rather than scalar measurements obtained by model fitting. This study demonstrates that microscopic tissue prediction based on pixel-by-pixel, and automatic model-free image segmentation based on diffusion-weighted imaging values ​​can be used in model training for healthy tissue and MS lesions. Diffusion peaks are widely known by 12 data points, bulging directions, and density estimates for only 8 data points. This provides a fast and robust method for clinical research, and also shows that standard data processing can be simplified with deep learning.
Discussion: Key issues and perspectivesMuch of the work already done shows that the use of deep networks has increased the current highest level, and these improvements are consistent in many areas. Often, the progress made by deep learning in giving solutions is relatively straightforward, and we can see this significant advancement in the field of medical computing. In the article "Deep Learning in Medical Imaging: A Review and Prospects of an Exciting New Technology," the question is raised: The large-scale species recognition theory in 2012 has a 10% improvement, but how to apply it. Has the aspect achieved a substantial leap? Is the question raised correctly? Is the direction of exploration correct? Is the image representation technique used enough to support it (eg 2D or 3D)? Need to get more data from each medical case? Or is it more efficient to turn to deep learning? There are more related issues raised in the second part of this article. Most of the problems need to be resolved.
In this document, it can be seen that although supervised learning and unsupervised learning are accessible through deep networks, it seems that most of the work is using supervised learning. So what about the medical field? The amount of data is a key factor, and the form requires the combination of both supervisory and unsupervised learning. In the medical field, it is difficult to obtain big data (manual annotation is difficult to obtain), so the field needs more semi-supervised learning and supervised learning.
This article includes a lot of network architecture. It can be seen from the current published paper that the variability is very large. Choosing a known architecture, designing a stable framework for tasks, and merging across frameworks can lead to variability. We can ask an interesting question about this: If a very deep residual network spans 152 layers and performs best on the ILSVRC 2015 layer classification task, can it be very good for medical applications?
An important aspect of deep learning is that it can benefit from a large amount of training data. After the ILSVRC competition based on the ImageNET dataset, I achieved a huge breakthrough in computer vision (). Compared to the training and test data sets used in other papers, the data set used for this particular problem is very large (compared to one million and one thousand one hundred). If we can build a public medical image data set like this, our society will benefit a lot.
Why is this job very challenging? First, it is difficult to raise funds for building such a data set. Second, high-quality annotation of medical image data requires medical expertise, which is not only scarce and very expensive; third, compared with natural images. Privacy issues make medical data more difficult to obtain; fourth, the breadth of medical imaging applications needs to collect more diverse data sets. Despite these many obstacles, we have made rapid progress in data collection and sharing of data. Many public data sets have been released and are now being used in practice. For example, VISCERAL and the Cancer Imaging Archive, Roth et al. [13] and Shin et al. [17], datasets obtained by analyzing CT scan-enlarged lymph node images have been published in cancer imaging archives, the same study The group also published the pancreatic dataset online.
Since 2007, it has become a habit to hold competition seminars at medical imaging conferences such as MICCAI, ISBI and SPIE medical imaging. There are a large number of data sets and ongoing research on the website (). Using these common benchmark data sets has a distinct advantage over using only public data sets: the competition provides a precise definition of the task to be solved, and one or more evaluation metrics have been defined to provide a fair evaluation criteria for various algorithms. Without such an evaluation criterion, even if each algorithm uses the same data set, it is difficult to compare different methods of the same problem. For example, three of the studies (Anthimopoulos et al. [16], Shin et al. [17] and van Tulder et al. [18]) used CT scan data sets of the same interstitial lung disease with medical annotations, but the results they reported It is different.
In this regard, a study on this issue (Setio et al. [12]) has seen initial results in a challenge to lung nodules. This challenge was co-organized by the IEEE and ISBI conferences using a public LIDC/IDRI data set, and the system proposed in this article can be directly compared to its alternatives.
Last year, there was a competition based on machine learning applications based on medical image analysis. Kaggle organized a competition to identify color images of the fundus of diabetic patients for a prize of $100,000. The 661 teams submitted the results and provided a total of 8,000 images. These data were used in a special study (van Grinsven et al. [24]). recent. The second session of the heart volume was measured by an MRI image, and the end of the medical image analysis competition for the ejection fraction image was completed. A total of 192 teams participated in the competition with a prize of $200,000. In both competitions, the best competitors are using convolutional neural networks. Among the better algorithms used, contestants using big data sets and deep learning have shown greater advantages, and we hope that this trend will continue. In this case, competition in the next series of worldwide on improving the accuracy of various imaging cancer screenings may attract the attention of relevant people.
Research by Albarqouni et al. shows that online platforms, such as those in competitions, can be used for a variety of purposes. They will promote new collaborations, form solutions, and gain access to large amounts of data through crowdsourcing.
ZGAR Aurora 3000 Puffs
ZGAR electronic cigarette uses high-tech R&D, food grade disposable pod device and high-quality raw material. All package designs are Original IP. Our designer team is from Hong Kong. We have very high requirements for product quality, flavors taste and packaging design. The E-liquid is imported, materials are food grade, and assembly plant is medical-grade dust-free workshops.
Our products include disposable e-cigarettes, rechargeable e-cigarettes, rechargreable disposable vape pen, and various of flavors of cigarette cartridges. From 600puffs to 5000puffs, ZGAR bar Disposable offer high-tech R&D, E-cigarette improves battery capacity, We offer various of flavors and support customization. And printing designs can be customized. We have our own professional team and competitive quotations for any OEM or ODM works.
We supply OEM rechargeable disposable vape pen,OEM disposable electronic cigarette,ODM disposable vape pen,ODM disposable electronic cigarette,OEM/ODM vape pen e-cigarette,OEM/ODM atomizer device.
Aurora 3000 Puffs,ZGAR Aurora 3000 Puffs Pod System Vape,ZGAR Aurora 3000 Puffs Pos Systems Touch Screen,ZGAR Aurora 3000 Puffs Disposable Vape Pod System,3000Puffs Pod Vape System
ZGAR INTERNATIONAL(HK)CO., LIMITED , https://www.szvape-pods.com