Data Acquisition and Automatic segmentation of LV
The Philips computerized tomography device was used to obtain the cardiac CT image for 17 male and female with A-fib in 50-62 age groups. This device captured ten sets of timed frames at the same location with various contrast agents, including a whole cardiac cycle. Each data set has 409 images with 512x512 pixel resolution. We used cardiac CT images with delayed enhancement as our dataset. A few layers were skipped because to their unimportance after we extracted the relevant dataset (149 CT pictures) that displays the LV portion. The study was conducted in conformity with the Institutional Review Board of National Yang-Ming University, Hospital, as was the informed consent process. We used Segment CT software to automatically segment the LV [5]. By inserting points during the reconstruction process, it can be used to locate LV on short-axis stacks. Using mathematical calculation, to find the possibilities of scar area in the dataset, we used standard deviation (SD) and average of the pixel value. This concept was derived from the literature, which states that if three SD values are greater than the average intensity value of a healthy myocardium region, scar tissue is present [5, 8,9]. We concentrated on determining the pixel value of the LV myocardium wall region because every pixel present on the surface as HU provides clues to help us plan our research. According to the literature, HU allows for a simple method of characterizing specific tissue. The density of the tissue is represented by HU, which is directly proportional to the degree of x-ray attenuation to every pixel in the CT scan image [6, 10, 11]. In this step, we performed automated localization and cropping of the selected myocardium wall that corresponds to the LV of the heart, as well as calculating SD and average value.
Implementation of morphological operations and Patch creation
In this step, we tried to use fundamental morphological operators like erosion, dilation, opening, and closing in order to get information [12]. We establish the criteria as follows: the pixel value that will be regarded as a contrast area in the specific dataset must be higher than two STD and one average value. We could finally see the contrast area clearly. We extracted 25 patches that contained the scar region on the LV and normal tissue/non scar tissue with 25x25 dimensions.
Radon descriptor and Texture feature extraction
The term "content-based image retrieval" (CBIR) refers to the process of looking up and examining an image's content. This content can be found using a variety of criteria, including color, shape, texture, and many others. CBIR is a crucial component of computer vision research, particularly in expert medical systems that look to human knowledge to use computers to solve the problem [13]. The Radon transform method can be applied to CBIR applications. The projection of the image in various directions is determined by the Radon transform, an integral transform [14]. The visual characteristics of medical images can be described using it [15, 16]. This problem can be resolved utilizing the Radon descriptor and content-based medical image retrieval (CBMIR). Patches of each kind were delivered after extraction in order to create the Radon images for each patient using the Radon descriptor. We used the MATLAB R2018a platform for this task in our research. Using the local binary pattern (LBP) approach for numerical data, we retrieved texture features from the Radon pictures. Approaches to face detection and pattern recognition frequently employ this technique. An image is transformed into an array or image of integer labels by the LBP operator [ 15, 16, 17, 18]. These labels characterize the image's appearance at small scales. The models of conventional machine learning methods were then fed the extracted numerical feature data.