Left ventricle analysis in echocardiographic images using transfer learning

Transfer Learning Video Classification of Preserved, Mid-Range, and Reduced Left Ventricular Ejection Fraction in Echocardiography

Identifying patients with left ventricular ejection fraction (EF), either reduced [EF < 40% (rEF)], mid-range [EF 40-50% (mEF)], or preserved [EF > 50% (pEF)], is considered of primary clinical importance. An end-to-end video classification using AutoML in Google Vertex AI was applied to echocardiographic recordings. Datasets balanced by majority undersampling, each corresponding to one out of three possible classifications, were obtained from the Standford EchoNet-Dynamic repository. A train-test split of 75/25 was applied. A binary video classification of rEF vs. not rEF demonstrated good performance (test dataset: ROC AUC score 0.939, accuracy 0.863, sensitivity 0.894, specificity 0.831, positive predicting value 0.842). A second binary classification of not pEF vs. pEF was slightly less performing (test dataset: ROC AUC score 0.917, accuracy 0.829, sensitivity 0.761, specificity 0.891, positive predicting value 0.888). A ternary classification was also explored, and lower performance was observed, mainly for the mEF class. A non-AutoML PyTorch implementation in open access confirmed the feasibility of our approach. With this proof of concept, end-to-end video classification based on transfer learning to categorize EF merits consideration for further evaluation in prospective clinical studies.

An Efficient Capsule-based Network for 2D Left Ventricle Segmentation in Echocardiography Images

The segmentation of cardiac chambers is essential for the clinical diagnosis and treatment of cardiovascular diseases. It is demonstrated that in cardiac disease, the left ventricle (LV) is extensively involved. Therefore, segmentation of the LV in echocardiographic images is critical for the precise evaluation of factors that influence cardiac function such as LV volume, ejection fraction, and LV mass. Although these measurements could be obtained by manual segmentation of the LV, it would be time-consuming and inaccurate because of the poor quality and low contrast of these images. Convolutional neural networks, commonly referred to as CNNs, have emerged as a highly favored deep learning technique for medical image segmentation. Despite their popularity, the pooling layers in CNNs ignore the spatial information and do not consider the part-whole hierarchy relationships. Furthermore, they require a large training dataset and a large number of parameters. Therefore, Capsule Networks are proposed to address the CNNs limitations. In this study, for the first time, an optimized capsule-based network for object segmentation called SegCaps is proposed to achieve accurate LV segmentation on echocardiography images applied to the CAMUS dataset. The result was compared against the standard 2D-UNet. The modified SegCaps and 2D-UNet achieved an average Dice similarity coefficient (DSC) of 84.48% and 83.28% on LV segmentation, respectively. The capabilities of the CapsNet led to an improvement of 1.44% in DSC with 92.77% fewer parameters than the U-Net. The results indicate that the proposed method leads to accurate and efficient LV segmentation.Clinical Relevance- From a clinical point of view, our findings lead to more precise evaluations of critical cardiac parameters, including ejection fraction as well as left ventricle volume at end-diastole and end-systole.