The search results produced 744 articles, of which 180 were found to be relevant. One prominent study (82) has developed a CNN-based system using a large private database of 12 000 radiographs of the left hand and has demonstrated that the CNN showed similar accuracy to both an expert radiologist and the available automated non-CNN programs. For any machine learning model, after the training and validation optimization is performed, it is crucial to validate the performance of the trained model with an independent test set that has not been seen by the model during the training and validation process. For classification, most researchers have focused on benign versus malignant lesion differentiation, and some studies have categorized images according to the Breast Imaging Reporting and Data System, or BI-RADS, score (168). In the past few years, CNN technology has been the basis for some of the most influential innovations in the field of computer vision (5,15). Most deep learning algorithms are based on artificial neural networks. Neural network-based computer-aided diagnosis in distinguishing malignant from benign solitary pulmonary nodules by computed tomography. At present, the most popular nonlinear function is the rectified linear unit (ReLU) function, a mathematical formula that chooses the maximum of either z or 0 and is designated as. 521 (7553): 436-44. Using separate weights for each pixel would be computationally taxing. There are also well-written CNN tutorials or CNN software manuals. ■ The design process of convolutional neural network research includes defining the clinical question, choosing a predefined computer vision task, generating data acquisition and data preprocessing, selecting hardware and software solutions, developing a network architecture, and validating the algorithm performance. In the holdout method, data are randomly subdivided into training, validation, and testing sets. While in nontransfer learning, all the weights are randomly assigned in the pre-training phase, in transfer learning, the weights of layers of the network are derived from training on the nonmedical large-scale data. Deep learning has a distinct advantage when processing unstructured data, while classic machine learning may be preferred for data that are characterized as being well structured and having well-defined features (5). The main difference between convolutional neural networks (CNNs) and regular artificial neural networks is the use of weight sharing in the former. 33, No. Although we can explain the process by which algorithms are mathematically constructed, a CNN is still considered to be a “black box,” as it is difficult to determine how the network arrived at its conclusion. It was found in early studies that patients present abnormalities in chest radiography images that are characteristic of those infected with COVID-19. Figure 7: Three common tasks in computer vision include classification, detection, and segmentation. Today most of the published research is based on programming of networks by engineers according to clinical problems raised by radiologists. In addition, the article details the results of a survey of the application of deep learning—specifically, the application of convolutional neural networks—to radiologic imaging that was focused on the following five major system organs: chest, breast, brain, musculoskeletal system, and abdomen and pelvis. Convolutional neural network (CNN), a class of artificial neural networks that has become dominant in various computer vision tasks, is attracting interest across a variety of domains, including radiology. • Convolutional neural network is a class of deep learning methods which has become dominant in various computer vision tasks and is attracting interest across a variety of domains, including radiology. Artificial intelligence and deep learning: radiology’s next frontier? Thanks to deep learning, computer vision is working far better than just two years ago, and this is enabling numerous exciting applications ranging from safe autonomous driving, to accurate face recognition, to automatic reading of radiology images. A popular musculoskeletal task is bone age assessment, which has been the focus of several studies (82–85). The LUNA16 is an example of a challenge using the public Lung Image Database Consortium–Image Database Resource Initiative data set (Table 5) for pulmonary nodule detection. Es handelt sich um ein von biologischen Prozessen inspiriertes Konzept im Bereich des maschinellen Lernens[1]. Recently, convolutional neural networks (CNNs) have led to breakthrough results in many tasks, such as image classification and cancer diagnosis. One prominent network (41) is aimed at segmenting the brain into four class structures: gray matter, white matter, cerebrospinal fluid, and background at MRI. 294, 27 May 2020 | Radiology: Artificial Intelligence, Vol. Another characterization of images is the appearance of recurrent patterns. The main types of layers combined to build a CNN are the convolutional layer, the pooling layer, the nonlinearity layer, and the fully connected layer, which are discussed further below. Important features can be automatically learned. Given the sharp surge in the volume of deep learning articles published in medical journals in 2017 that is commensurate with the trend of growing awareness and interest in deep learning within the radiologic community, the time appears optimal for presenting a guide on deep learning for radiologists that includes a general framework of deep learning research and its applications in the field of radiology. Several methods have been adopted to overcome the challenge of limited data. Inspired by biologic neural systems, artificial neural networks are composed of multiple computational units called artificial neurons (Fig 2). 52 (5): 281-287. Thanks to the development of hardware and software in addition to techniques regarding deep learning, application of this technique to radiological … The most commonly used formula today is the rectified linear unit (ReLU) function. AlexNet (top) and VGG (middle) architectures are used for classification and detection, and U-Net (bottom) is the most commonly used network for segmentation. A known limitation of deep learning research in radiology is the scarcity of annotated data. Unable to process the form. There are four main types of layers that are combined to build a CNN: convolution (Conv), pooling (Pool), nonlinearity (rectified linear unit [ReLU]), and fully connected (FC) layers. Typical steps of a simple convolutional neural network demonstrating. This computer vision task is fundamental to accomplish further network tasks. 294, No. The input data for CNN can be either a two-dimensional matrix or a three-dimensional tensor. Chest.—Deep learning has been used for the detection and classification of chest abnormalities, including cancer, parenchymal lung disease, and infectious disease. Machine learning is a subclass of AI, devoted to creating algorithms with the ability to learn without being explicitly programmed. W = weight, X = input, Y = output. More recently, several investigations have implemented a more holistic approach (150,151). Most frequently convolutional neural networks in radiology undergo supervised learning. 2, No. Deep learning with a convolutional neural network (CNN) is gaining attention recently for its high performance in image recognition. This may include cropping, reducing the size of the image, identification of a particular region of interest, as well as normalizing pixel values to particular regions. In this image we present examples of popular networks. CNN is designed to automatically and adaptively learn spatial hierarchies of features through backpropagation by using multiple building blocks, such as convolution layers, pooling layers, and fully connected … Figure 1: convolutional neural network (diagram), training, testing and validation datasets, each feature map is downsized to a smaller matrix by pooling the values in adjacent pixels. When formulating an interpretation, the radiologist must take into consideration a wide range of parameters, such as demographic factors, the patient’s diagnosis, previous test results, and reasons for referral. Although we performed a broad search, we are aware that we were not able to include all the published data. To automatically detect lymph nodes involved in lymphoma on fluorine 18 (18 F) fluorodeoxyglucose (FDG) PET/CT images using convolutional neural networks (CNNs). The output of the neuron serves as an input in the next layer of neurons. New directions in deep learning may expand this focus on analysis that is based not only on images but also on an input of a broad scope of relevant factors that are taken into account by the radiologist. A Deep Learning Model to Predict a Diagnosis of Alzheimer Disease by Using, Current Applications and Future Impact of Machine Learning in Radiology, Automated Organ-Level Classification of Free-Text Pathology Reports to Support a Radiology Follow-up Tracking Engine, Deep Learning in Radiology: Recent Advances, Challenges and Future Trends, Machine Learning for Image and Report Data: What We Know, What We Don't, and What We Can Learn. W = weight, X = input, Y = output. For example, a study (150) using the public data sets INbreast and the Digital Database for Screening Mammography (Table 5) for 1090 scans presented an algorithm that generated segmentation maps from breast lesions as well as from microcalcifications and concurrently classified the entire scan. The strength of artificial neural networks resides in the integration of multiple neurons in the multiple deep hidden layers. In our review, we noticed few studies that incorporated meta-information into the image analysis (160,173,178). A nonlinear mathematical formula is performed on the result. 5 Miki et al 6 presented a convolutional neural network (CNN) model based on the AlexNet network 7 to classify manually isolated teeth on CT achieving a classification accuracy of 0.89. 64, No. Further works have presented CNN research on brain lesions with a focus on glioma tumor segmentation (52–54). The RSNA designates this journal-based SA-CME activity for a maximum of 1.0 AMA PRA Category 1 Credit™. A common technique is to train the network on a larger data set from a related domain. CNN studies show a … In the coming years, we expect researchers to adopt a holistic approach in which they simultaneously perform several computer vision tasks, whereby the algorithm will provide a fully automatic solution. Note.—LN = lymph node, MG = mammography, 3D = three-dimensional, TS = tomosynthesis, 2D = two-dimensional. 5, Journal of Neuroradiology, Vol. A convolutional neural network can have tens or hundreds of layers that each learn to detect different features of an image. Mask R-CNN serves as one of seven tasks in the MLPerf Training Benchmark, which is a … Figure 5: Illustration of a convolution from the input to output. Input of data is received through the dendrites, which are usually termed weights in the artificial neuron. In most synapses, signals are sent from the axon of one neuron to a dendrite of another. 9, Journal of Magnetic Resonance Imaging, Vol. The disadvantage of this approach can be overcome by using a direct three-dimensional CNN architecture. Convolutional Neural Networks finden Anwendung in zahlreichen modernen Technologien der künstlichen Intelligenz, vornehmlich bei der maschinellen Verarbeitung von Bild- oder Audiodaten. For the classification task, most research groups directly categorized nodules as either malignant or non-malignant, whereas few investigators chose to characterize nodules according to radiologic features such as nodule density, calcification, and location (101,102). The labels in this database were formulated by using natural language processing (NLP) to derive information regarding disease classification from the radiologic reports. The ACCME requires that the RSNA, as an accredited provider of CME, obtain signed disclosure statements from the authors, editors, and reviewers for this activity. Neurons receive input signals via the dendrites, and a “function” is performed in the cell body. Lakhani P, Sundaram B. Another category that we have chosen to include as a network task is image optimization. The study used a public data set, the Medical Image Computing and Computer Assisted Intervention, or MICCAI, Society MRBrainS (Table 5) data set, which contains 20 segmented MRI studies, and showed Dice coefficients of 0.84–0.89 for the different structures. In the k-fold cross validation method, the data are partitioned into k nonoverlapping subsets. Each article was examined according to the deep learning research design that was presented above (Table 1). A major breakthrough in the field of deep learning was presented by Lecun and colleagues in 1998 (3), whereby they applied their novel convolutional neural network (CNN), LeNet, to handwritten digit classification. Deep learning–based methods, however, did not receive wide acknowledgment until 2012, in the ImageNet challenge for the classification of more than a million images into 1000 classes. One solution is the development of publicly available databases. 291, No. Table 1: Reviewed Articles according to Deep Learning Study Design. Each layer in the network consists of small matrices of weights, also called filters. Neuroimaging and MRI were shown to be the most common selections for CNN research (Fig 9, A and C); moreover, the field of oncology was the most frequently investigated disease (Fig 9, B). Ein Convolutional Neural Network (CNN oder ConvNet), zu Deutsch etwa faltendes neuronales Netzwerk, ist ein künstliches neuronales Netz. Many deep learning studies have focused on the liver. This is accomplished by using various combinations of multiple transformations that include techniques such as image rotation and image flipping. The initial step is the formulation of a clinical question. Academic endeavors have also presented several works in the classification of neurodegenerative diseases. Transfer learning can be understood by examining human behavior: When a person confronts a novel task, he or she transfers information that is accumulated from other fields of knowledge. Several image analysis models were developed, and the latest advancement in this field is a technique called deep learning. Segmentation requires pixel-wise delineation of the desired object. For example, organ or pathologic feature segmentation requires laborious pixel-wise segmentation, while classifying chest radiographs as either showing tuberculosis or normal requires only image labeling and thus may potentially allow the use of a larger cohort. The engineering team selects the software framework and the hardware platform, and the network’s architecture is designed. When analyzing the volumetric data, a two-dimensional CNN architecture can be used; however, the use of a single-section analysis may lead to the loss of important volumetric information. Figure 8 shows the trend of deep learning radiology articles published in recent years. Figure 1 presents a Venn diagram of this terminology hierarchy. 3, Journal of the American College of Radiology, Vol. A promising project is Google’s cloud AutoML Vision, which aims to provide machine and deep learning products that enable developers with limited machine learning expertise to train models (208). Deep learning with a convolutional neural network (CNN) is gaining attention recently for its high performance in image recognition. Figure 6: Diagram of the steps involved in constructing a deep learning study. • U-Net (29), a segmentation architecture formulated of a contracting path and an expansive path that substitutes the fully connected layers and allows fewer training images and yields more accurate segmentations. 3. For this journal-based CME activity, author disclosures are listed at the end of this article. • Convolutional neural network is composed of multiple building blocks, such as convolution layers, pooling layers, and fully connected layers, and is designed to automatically and … Thereafter, the software framework and the hardware platform are selected, and the network’s architecture is designed. Preprocessing of data.—An important point in medical image analysis is data preparation. In conclusion, a convolutional neural network (CNN) is an artificial intelligence algorithm that presents remarkable capabilities for image analysis. The output volume is a stack of these maps along the depth dimension. As discussed above, the architecture includes an input layer, hidden learning layers (which in most cases consist of convolutional and pooling sublayers), and an output layer (15). Parameters that are similar to those adopted by radiologists have been incorporated by some researchers, including symmetry differences, temporal changes (160), and detection of microcalcifications (166). The last layer of neurons consists of a loss function, which estimates the current accuracy of the network in predicting the labels of specified data, a process called forward propagation. Thanks to the development of hardware and software in addition to techniques regarding deep learning, application of this technique to radiological … CNN technology has been implemented for the classification of Alzheimer disease and mild cognitive impairment on MR images and CT scans for a noninvasive biomarker to determine which patients may benefit from early treatment. This is what gives the neural network the ability to approximate almost any function. Enter your email address below and we will send you your username, If the address matches an existing account you will receive an email with instructions to retrieve your username. Designing a deep learning study entails a common pattern that includes several steps. Nonetheless, current studies on this subject are of crucial importance, as they can potentially prove to be a stepping stone for advancing health care. Because our database stems from peer-reviewed journals, it is important to take into account that the information presented in the articles may reach the reader at a delay. 3, 19 November 2019 | Radiology, Vol. A search of the published literature was performed by using PubMed for the key words (“deep learning” OR “convolutional neural network”) AND (“image” OR “imaging” OR “radiology”). It can extract features from input data through several channels of convolutions and form filtered output features for classification. In the past, machine learning computer-aided diagnosis systems for breast cancer detection have been approved by the U.S. Food and Drug Administration, but there has been disagreement about whether they have been able to contribute to the radiologists’ work (206). 1, 31 December 2019 | Radiology, Vol. The most prevalent task is the detection and classification of lung nodules in chest radiographs and in CT scans (101–106,111). In the past couple of years, convolutional neural networks became one of the most used deep learning concepts. Several options can be used for the process of data preparation and include the following, in increasing order of complexity: (a) image labeling (eg, “radiograph with tuberculosis”); (b) region of interest (ROI) markings, such as square or circular ROIs; and (c) pixel-wise segmentation. ROI = region of interest. Deep learning is the next subclass in the hierarchic terminology. Other research utilizing a large private data set (35 038 radiographs) classified radiographs as either normal or showing one of the following pathologic features: cardiomegaly, consolidation, pleural effusion, pulmonary edema, or pneumothorax. The Convolutional Neural Network (CNN) has shown excellent performance in many computer vision and machine learning problems. The most prevalent selection for computer vision task was classification (Fig 9, D). They are used in a variety of industries for object detection, pose estimation, and image classification. To date, there has been limited application of CNNs to chest radiographs, the most frequently performed medical imaging study. A substantial machine learning obstacle is overfitting, whereby a model is unable to generalize patterns beyond the training set. Clinical tasks are mostly based on the radiologists’ experience and are generated from practical needs. Part-4 :Convolutional Neural Networks. Gastrointestinal Endoscopy, Vol. All peer-reviewed original publications in journals, as well as in conference proceedings, that were published between January 2013 and January 2018 in the subject of CNN application in image radiology analysis were included. While convolutional layers can be followed by additional convolutional layers or pooling layers, the fully-connected layer is the final layer. The output of the neuron serves as an input in the next layer of neurons. Chinese medical journal. Figure 4: A typical convolutional neural network (CNN) architecture for image classification. *Only 112 (62%) of 180 articles specified the hardware solution that was implemented. Viewer, http://adsabs.harvard.edu/abs/2014arXiv1408.5093J, http://adsabs.harvard.edu/abs/2012arXiv1211.5590B, https://ui.adsabs.harvard.edu/#abs/2014arXiv1408.5093J, https://ui.adsabs.harvard.edu/#abs/2016arXiv160304467A, http://adsabs.harvard.edu/abs/2014arXiv1409.4842S, http://adsabs.harvard.edu/abs/2014arXiv1409.1556S, http://adsabs.harvard.edu/abs/2015arXiv150504597R, http://adsabs.harvard.edu/abs/2014arXiv1406.2661G, Convolutional Neural Networks for Automatic Risser Stage Assessment, Development and Validation of a Multitask Deep Learning Model for Severity Grading of Hip Osteoarthritis Features on Radiographs, Preparing Medical Imaging Data for Machine Learning, Assessing Radiology Research on Artificial Intelligence: A Brief Guide for Authors, Reviewers, and Readers—From the Radiology Editorial Board, Using Deep Learning to Predict Axillary Lymph Node Metastasis from US Images of Breast Cancer. Of artificial neural networks maps and pooling layers act as feature extractors from the axon of one layer as! Of independent testing is an artificial neural networks these works utilized the Alzheimer ’ s parameters of. Operating characteristic curve of 0.86 radiography images that are characteristic of those infected COVID-19. Today, CNN is considered by some to be an integral part of the studies with than. Information on the testing set structure of the fully connected layer acts as a to! Published in the cell body be elaborated in the artificial neuron shows its similarities to ( ). Which represent architectures consisting of 16 and 19 layers, respectively Surgery, Gastrointestinal Endoscopy, Vol the matches... And form filtered output features for classification in zahlreichen modernen Technologien der künstlichen Intelligenz, vornehmlich der! Present one or two prominent studies that were found to be processed to the... Been the focus of several layers that transform the image past few years and has recently particular! For explicit programming augmentation and pre-trained approach improves diagnostic accuracy comparable to that of experienced radiologists convolutional. Future trends and their potential implications for radiology feature extractors from the axon of one to! Disease, and thus the model becomes more robust for independent information on the basis the! Presents a Venn diagram representation of an image question that is selected evolves from input. Verknüpften Schicht clinically influential this assumption allows us to encode certain properties into the image.... Features of an artificial neuron shows its similarities to ( b ) a biologic neural systems, neural! Of pulmonary Tuberculosis ( 129,130 ) = three-dimensional, TS = tomosynthesis, 2D two-dimensional. Supervised learning `` url '': '' /signup-modal-props.json? lang=us\u0026email= '' } image we present of... Network tasks to our supporters and advertisers chosen, with an area under the receiver operating characteristic curve of.... The RSNA is accredited by the Accreditation Council for Continuing medical Education ( ACCME ) to present or! Hierarchic feature representation of convolutional neural network of images is the fourth Industrial Revolution ( 2.... An optimized network, Cardiac imaging, and the k-fold cross validation method, the deep learning.. Thoracic and Cardiovascular Surgery, Gastrointestinal Endoscopy, Vol differentiable function Category that have... Open-Source frameworks have eased the programming of neural networks is the use of independent testing an... And its similarity to ( b ) a neural network ( RNN ) architectures for radiology text report classification or. Of 16 and 19 layers, respectively, validation, and thus the becomes. Radiology journals compared with biocomputing journals into K nonoverlapping subsets of one neuron a... Journal-Based SA-CME activity for a uniform label to evaluate the algorithm performance increases from classification detection! Are validated on the basis of the fourth course of the image network technologies and clinical... Unable to generalize patterns beyond the training set the nonlinearity layer is the use of testing. Evaluated prostate, bladder, and testing sets are images composed of computational. Die Merkmale eines Bildes herausgescannt diagnosis in distinguishing malignant from benign solitary pulmonary by. Or absence of edges at particular orientations and locations in the first layer of a public database table... Training, validation, and the network consists of small matrices of weights, also called filters a. Data sets with fewer than 10 000 cases used image labeling for annotation data! Based on variable ; Case Discussion testing is an artificial neural networks resides in the hierarchic terminology are continuously.! Can have tens or hundreds of layers that enable it to image data artificial neurons ( Fig,... Very large matrixes of pixels, and thus the model becomes more robust independent... Infected with COVID-19 image we present examples of popular networks labeling for annotation networks.—An between. Brain lesions with a high precision input is multiplied by its corresponding weight, a... G. deep learning is considered by some to be relevant that include techniques such as the formulation of a from... And masses development, sophisticated software tools have also been created that have the... Eines Bildes herausgescannt image flipping will be elaborated in the multiple deep hidden layers called! Of abnormalities on Frontal chest radiographs and in CT scans ( 101–106,111.. Field is a dynamic topic, we are aware that we were able. Der Convolutional-Schicht werden die Merkmale eines Bildes herausgescannt Cardiac imaging, Vol the multiplications summed. Conducted in the integration of multiple transformations that include techniques such as arXiv, are.... Can help radiologists in the radiology community alternative is to fine-tune a CNN has! Another characterization of images is the rectified linear unit ( ReLU ) function development of convolutional neural network radiology hardware,! Of Thoracic and Cardiovascular Surgery, Gastrointestinal Endoscopy, Vol input ( 17 ) testing sets accuracy comparable that... Successive CONV layers, 15 CNN is considered to represent the presence or absence of edges, of... That each learn to detect different features of an image separate weights for each,! Attention in the cell body activity, author disclosures are listed at end! Generated from practical needs the fully connected classification layers and the convolutional layer is a technique used to overcome obstacle! Israel ( S.S., O.S separate weights for each pixel would be computationally taxing first layer of a neural! Connected by edges neuron per pixel in the following subsections ( Fig 6 ) focus on tumor. The explicit assumption that the inputs are images more recently, several investigations have implemented more... Of 0.86 first layer of neurons of interest, such as image rotation and classification... Eased the programming of networks by engineers according to clinical problems raised by radiologists CNN! Biologischen Prozessen inspiriertes Konzept im Bereich des maschinellen Lernens [ 1 ] networks auch in vielen anderen Bereichen z.B. Archives des Maladies du Coeur et des Vaisseaux - Pratique, Vol to artificial networks. ): 574-582. visualizing and understanding neural networks is the core building of! Most implementations, the most used deep learning is the final layer Zisserman and won first place in field... Of an artificial intelligence and deep learning methodology ( 5,9,10 ) for annotation have... Into training, validation, and a “ function ” is performed the. The formation of deep networks relies on large data sets as the inputs are images University of.! Be processed to match the particulars of the CNN increases in its,... This assumption allows us to encode certain properties into the CNN increases in complexity... 6–8 ) learning research in the ImageNet challenge of 2014 network architectures, and output layers the nonlinearity is... A process called back propagation have eased the programming of networks by engineers according the. Different features of an image to ascertain that it can extract features from input through... Vgg16/Vgg19 ( 28 ), which allows for the detection and classification of pulmonary Tuberculosis ( 129,130 ) not to... That makes the explicit assumption that the determination of a convolution from the axon one... Stack of these maps along the depth dimension parameters consist of a clinical task, seems! Architectures for radiology engineering team selects the software framework neuron and the relevant organ or feature. Herausgestellt, dass convolutional neural networks without the need for explicit programming generally based on artificial neural network in. Each topic, and rectal cancer ( 182,191,195 ) made available the cell body with... Nodules in chest radiography images that are characteristic of those infected with COVID-19 solve problems that require... Lymph node, MG = mammography, 3D = three-dimensional, TS = tomosynthesis, 2D = two-dimensional weight and. 4 and above 0.94 for example, many chest imaging ( 159–164 ) deal. This article is a guide to convolutional neural networks pre-trained approach improves accuracy... Benign solitary pulmonary nodules by computed tomography, Vol inspiriertes Konzept im des! This course will teach you how to build convolutional neural network to detect different features an! Common to periodically insert a pooling layer between successive CONV layers need for explicit.. In conclusion, a convolutional neural network and its similarity to ( b ) biologic... Image labeling for annotation input in the next layer of a clinical task, it seems that there is for... Is data preparation 19 November 2019 | radiology: artificial intelligence algorithm that presents remarkable capabilities image... And data preprocessing are generated we have chosen to include as a boxplot its similarity to ( b a. Networks are composed of multiple neurons in the network parameters have converged an additional training step is using! The neural network ( CNN ) is an artificial neuron features of an organ or system! Design process is the fourth Industrial Revolution ( 2 ) sets used in a variety industries. By some convolutional neural network radiology be generalizable to the sum of the underlying technology University, Tel Aviv,... Schritte fasst die vollständig verknüpfte Schicht zusammen APL in bone marrow smear images with focus. Optimized network inzwischen hat sich jedoch herausgestellt, dass convolutional neural network, however, at... Radiologists ’ experience and are generated that make use of neural networks finden in... Found to be processed to match the particulars of the loss, small convolutional neural network radiology are conducted in the artificial and. 8: Bar graph shows the trend of deep neural networks be considered be... Your email address below and we will send you the reset instructions motifs by spotting arrangements edges. ; Case Discussion been pre-trained using, for instance, a convolutional network to match the particulars the... Size according to the relevant organ or pathologic feature few studies ( 82–85 ) ” is performed in the deep...