Deep Learning In Echocardiography: A Review
Abstract
Artificial intelligence extending the horizon of all domains of science, technology and even our day today life, deep neural network is a sub domain of ANN which is also the subdomain of artificial intelligence. This review article will explain the role and current application, researches of deep neural network in echocardiography and also discuss the limitations and challenges in it. Echocardiography is widely used to determine the various heart diseases, functions and also anatomy. Analysis and interpretation of echocardiography requires expert knowledge, due to this it is costly, time consuming and critical to take decisions. Interpretation of cardiac image using automated computer systems have drastically transform the clinical practices by identifying the abnormalities in heart muscles motions which is used to determine the heart disease. Deep learning technique is use to analysis images, now it is used in medical imaging problem and also very useful for physician for enhancement patients care. Unlike statistical approach it requires large dataset of images for training the model for specific problem. When deep learning is applied to large datasets, it determines and establishes the complex patterns and their relationship in image. Computer learn through large dataset and recognises the desire pattern in image. Although automated systems in medical science are not widely accepted, this technique not only help practitioners but also academician. Keywords- deep learning, echocardiography, heart disease, automation.
Introduction
Recent advancements in technologies of computer science both in hardware’s and software’s have a great impact in every domain of science and technology are significantly influenced and used by artificial intelligence(AI).
In other words, AI is the ability of a machine to perceive its environment and perform measured action to maximize its chance of success for a specified goal1. In AI we try to mimic the human intelligence in computer system or roughly in a machine with different mechanism which are used by human brain in decision making or problem solving process of a particular problem. Our thinking process is very much complex, it all govern by our brain cells or nerve cells, to build a copy of brain is very difficult, so researchers try to copy the one or few behaviours of them to solve the same kind of problem as they are used in human brain. Our brain evolved in thousands of years, we used to learn so many things since our childhood, our thinking process is not based upon one aspect of problem, it based upon various condition and situations, so taking decision or solve a problem by human beings is not difficult and also learning new things which have connections what we have learned in past, is easy. Our modern computing devices are very fast, applying artificial neural network which is the mimic of biological neuron which is highly parallel, robust and scalable can solve many complex tasks such as weather forecasting, stock market prediction very quickly and with high accuracy.
Deep neural network is the sub-domain of artificial neural network which is also comes under the umbrella of machine learning and its process of learning is called deep leaning, it is mainly used in image analysis. Heart diseases which are associated with vascular problem requires to view the live internal structure of heart while its normal movement(beat), so that practitioners can identify the disease and also severity of disease which help him to enhance the care of patient, but it requires expert knowledge in this domain. With the help of echocardiography one can look inside the heart when it is pumping blood to the body. In echocardiography laboratory, experts rigorously analyse and label thousands of echocardiography images and videos every day which help us to learn and identify different heart diseases. Images are very complex in nature they have various information in them which is known as pattern which make them different from other images. This conclusion is result of Conscious and sub-Conscious state of mind and it is also takes years of dedication and experiences. When we see an object we identify its pattern, next time whenever we see other object which has similar pattern as we saw previously we draw a conclusion that, this is that type of object. Deep learning does the same thing, it input as a large dataset of images, identify patterns in them, it is different from statistical approach in which we estimate values of patterns.
Artificial Intelligence, Machine Learning, Deep Learning
Advancement in technology in medical science whether in medicine, surgery or disease diagnosis expanding the boundaries of our thinking to make better decision support system, in other word expert system which can help in various way. Machine learning is a component of AI described as the process for a computer to learn from experiences and perform predefined tasks without prior knowledge3. Machine learning techniques further classified in supervised learning, Semi supervised learning, and unsupervised learning which is based on what kind of data used for learning such as fully labelled, partially labelled, or unlabelled. One of the example of machine learning technique is an artificial neural network. An artificial neural network is a collection of layers of neurons and they are connected to each other according to their tasks, they have their own weights to reflect their inter-dependencies between other neurons or same neurons themselves which is more likely to be a human brain. In this network each neuron receives multiple inputs, weight and bias, and combine them together in some mathematical way, coming towards all inputs determine the activation or firing of that particular neuron.
Activated neurons give an output signal which determine the states of neurons in the very next layer, until it reaches the neuron of final layer which provides values of our desire output such as a classification, decision making or estimation of value (figure 1). Learning is a process in which weights are adjusted between interconnection of neuron by providing data samples, to get desire output. As the number of neurons increase and grow of sample size, complexity of learning process also increases. Due to limitation of computing power and increasing size of samples, the success rate of a machine learning algorithm depends upon features extraction from raw samples to reduce to number of neurons which ultimately reduce the complexity and cost. Deep learning was come to exist to overcome this limitation by learning automatically from representing features in samples. Extracting optimal features, deep learning used a number of cascaded multiple layers of neurons for learning of multiple levels of abstraction. In well-designed neuron network structures, deep learning is able to do the same task with good performance such as image recognition, speech recognition and prediction of activity of drug molecules. The motive behind the research was to identify whether an image is recoded from an apical four-chamber view, a parasternal short-axis view, or a parasternal long-axis view. The input of the Convolutional layer is a fairly taken standard echocardiographic image, and the output is the probabilities of the different outcomes, which are used to determine the most similar view. During deep learning process images are analysed through multiple layers of abstraction allows the computer to identify the view automatically from the image taken from7. In this convolutional neural network example, features are learnt from data by it in the training phase (all echocardiographic image dataset) instead of manual extraction. Limitation of this method is, it is time consuming due to analysis of large dataset also very complex. The used classifier is softmax (pink circles) having up to 15 nodes b. the whole dataset of images was split into three-part training, validation and testing and test data was not used for validation and training.
In the domain of medical imaging and diseases diagnosis accountability and responsibility play a very important role, since medical imaging is opening the door to the additional invasive diagnostic treatments and testing, which may uncover a patient to excessive risk. Definitely, in current medical practices, doctors will ultimately must need to take responsibility for interpretation of output comes from deep learning model. Even if there are limitations in deep learning is impulsive force in medical imaging technique and will be a most important contribution of the future of echocardiography imaging technique. As a researcher we will have to build or modelled some deep learning model which push the horizon of technologic discovery.
