Assessing left ventricular ejection fraction for cta imaging with ai-powered algorithms

OBJECTIVE Cardiac computed tomography (CT) is a diagnostic tool for detecting heart diseases, often evaluating cardiac function using left ventricular ejection fraction (LVEF) on multi-dimensional computed tomography (MDCT). However, inter-observer and intra-observer biases arise due to hand-drawn regions of interest (ROI) on images. This study aims to utilize fully convolutional network (FCN) models to segment the volume between systolic and diastolic phases, reducing biases. MATERIALS & METHODS Ten phases of MDCT were involved, with 53 2D images per phase, resulting in a total of 530 images with a size of 512x512 pixels. Ground truth boundaries were hand-drawn for each 2D image. Three optimizers, including ADAM and SGDM, were applied. FCN models were built for LV segmentation on MDCT using ResNet50, ResNet18, Mobilenetv2, Xception, and Inceptionresnetv2. The hyperparameters included batch size, epoch size, and learning rates (0.001). The dataset was divided into 70%, 20%, and 10% for training, validation, and testing of the FCN models, respectively. The performance index was evaluated using LVEF rate, IoU (intersection of union), and dice score among FCN models. RESULTS The ResNet50 model with hyper-parameters of SGDM optimizer, batch size 3, and epoch size 9 provided the best LVEF rate (0.59), IoU (0.80), and dice score (0.71). CONCLUSION The study demonstrated that FCN is an efficient, automatic, and Objective: segmentation method for LV on MDCT, making it a viable AI method for delineating cardiac boundaries of LV. Cardiac computed tomography (CT) is a diagnostic tool for detecting heart diseases, often evaluating cardiac function using left ventricular ejection fraction (LVEF) on multi-dimensional computed tomography (MDCT). However, inter-observer and intra-observer biases arise due to hand-drawn regions of interest (ROI) on images. This study aims to utilize fully convolutional network (FCN) models to segment the volume between systolic and diastolic phases, reducing biases. Ten phases of MDCT were involved, with 53 2D images per phase, resulting in a total of 530 images with a size of 512x512 pixels. Ground truth boundaries were hand-drawn for each 2D image. Three optimizers, including ADAM and SGDM, were applied. FCN models were built for LV segmentation on MDCT using ResNet50, ResNet18, Mobilenetv2, Xception, and Inceptionresnetv2. The hyperparameters included batch size, epoch size, and learning rates (0.001). The dataset was divided into 70%, 20%, and 10% for training, validation, and testing of the FCN models, respectively. The performance index was evaluated using LVEF rate, IoU (intersection of union), and dice score among FCN models. The ResNet50 model with hyper-parameters of SGDM optimizer, batch size 3, and epoch size 9 provided the best LVEF rate (0.59), IoU (0.80), and dice score (0.71). The study demonstrated that FCN is an efficient, automatic, and Objective: segmentation method for LV on MDCT, making it a viable AI method for delineating cardiac boundaries of LV.