Heart Failure is a major component of healthcare expenditure and a leading cause of mortality worldwide. Despite higher inter-rater variability, endomyocardial biopsy (EMB) is still regarded as the standard technique, used to identify the cause (e.g. ischemic or non-ischemic cardiomyopathy, coronary artery disease, myocardial infarction etc.) of unexplained heart failure. In this paper, we focus on identifying cardiomyopathy as ischemic or non-ischemic. For this, we propose and implement a new unified architecture comprising CNN (inception-V3 model) and bidirectional LSTM (BiLSTM) with self-attention mechanism to predict the ischemic or non-ischemic to classify cardiomyopathy using histopathological images. The proposed model is based on self-attention that implicitly focuses on the information outputted from the hidden layers of BiLSTM. Through our results we demonstrate that this framework carries a high learning capacity and is able to improve the classification performance.
Deep Dive into Self-attention based BiLSTM-CNN classifier for the prediction of ischemic and non-ischemic cardiomyopathy.
Heart Failure is a major component of healthcare expenditure and a leading cause of mortality worldwide. Despite higher inter-rater variability, endomyocardial biopsy (EMB) is still regarded as the standard technique, used to identify the cause (e.g. ischemic or non-ischemic cardiomyopathy, coronary artery disease, myocardial infarction etc.) of unexplained heart failure. In this paper, we focus on identifying cardiomyopathy as ischemic or non-ischemic. For this, we propose and implement a new unified architecture comprising CNN (inception-V3 model) and bidirectional LSTM (BiLSTM) with self-attention mechanism to predict the ischemic or non-ischemic to classify cardiomyopathy using histopathological images. The proposed model is based on self-attention that implicitly focuses on the information outputted from the hidden layers of BiLSTM. Through our results we demonstrate that this framework carries a high learning capacity and is able to improve the classification performance.
Cardiovascular disease (CD) is a major cause of heart failures and related causalities worldwide. According to the survey done by Centres for Disease Control (CDC) in 2011, CD is the foremost cause of deaths in the United States, Australia, United Kingdom and Canada [1]. Heart failure is a severe, progressive clinical syndrome that results in inadequate systemic perfusion. It is evident through other common symptoms like arrhythmia (irregular heartbeats), myocardial infarction (commonly known as heart-attack), and coronary heart disease [2].
Cardiomyopathy is responsible for approximately one third of all clinical heart failure cases. It is a disease related to hardening of the heart muscle that leads to heart failure in several cases [3]. Depending upon the underlying cause, cardiomyopathy has been further divided into two categories: ischemic and non-ischemic cardiomyopathy [4]. Ischemic and idiopathic dilated cardiomyopathy (a type of non-ischemic cardiomyopathy) are the two most common causes of heart failure with left ventricular systolic dysfunction, the only definite treatment being the heart transplant. Ischemic cardiomyopathy is a coronary heart disease caused by left ventricle dysfunction due to chronic absence of oxygen to the heart muscle. On the other hand, nonischemic cardiomyopathy is not caused by coronary artery disease and is often associated with organ illnesses and exhibits common symptoms such as; breathlessness, sweating, fast breathing, high level of fatigue etc. [5]. The gold standard method for the differentiation between ischemic and non-ischemic cardiomyopathy is coronary angiography [6]. However, due to its high cost and invasive nature, it is not feasible to analyze all patients with systolic heart failure by coronary angiogram. In order to prevent unnecessary coronary angiography, it would be very useful to be able to differentiate patients suffering from non-ischemic cardiomyopathy non-invasively with adequate precision. Non-invasive techniques such as thallium scintigraphy are costly, while dobutamine-stressed echocardiography is techniciandependent and is not accessible in general [6][7][8]. The other non-invasive imaging modalities like compute tomography (CT), magnetic resonance imaging (MRI) are preferred choices because of improved image quality and diagnostic accuracy [7][8][9][10][11][12]. In addition to CT and MRI, an endomyocardial biopsy (EMB) is considered in some instances complementary; or is the only procedure for diagnosing both ischemic and non-ischemic conditions in unexplained cardiomyopathy cases [13][14][15]. The EMB based diagnostics, depends upon the conviction of pathologists and hence there is an ample room for inter-rater variability [16]. The quantitative interpretation of pathology images is very important for an accurate diagnosis [17]. At present, quantitative analysis in a standardized clinical environment is a time-consuming affair and carries high inter-rater variability, and so an automated quantification algorithm fostering diagnostic outcomes with universal acceptability is urgently required.
Recent, burgeoning and successful applications of deep learning in the field of medicine and digital pathology demonstrate its effectiveness in learning hidden patterns that may not be visible by human examination [16]. In fact, deep learning architectures have been shown to be successful in the automated classification and segmentation of disease in digital histopathology [17]. The benefit of models based on deep learning is that they become familiar with the most appropriate representation progressively, as part of the training process. However, in the traditional CNN’s output layer is fully connected to the hidden layer, but it includes inefficient or multiple kernel that extract the same or trivial information repeatedly from input data [17]. To improve the accuracy further, network can be enlarged by adding more kernels, convolutional layers, and pooling layers, but it will increase the computational cost and also there is a chance of overfitting [18,19]. LSTM is more appropriate for handling timedependence problems [20]. It can filter and fuse the empty input, similar information, and irrelevant information extracted from the convolutional kernels, so that the effective and relevant information can be stored for a long time in the state cell. Benefit of a memory cell and controlling gates is that it prevents the gradient from vanishing too quickly, thereby propagating information without loss [21][22][23]. But LSTM only exploits the preceding or past information. BiLSTM is an advancement of LSTM in which forward hidden layer is combined with backward hidden layer, that can access both the preceding and subsequent information.
The vector representation of high intra-variability and sparsity causes histopathological images of high-dimensional vector [24]. The high-dimensional vector acting as the BiLSTM input will raise the number of network para
…(Full text truncated)…
This content is AI-processed based on ArXiv data.
