Evaluación temprana de la lesión miocárdica en pacientes con enfermedad por coronavirus 2019 utilizando un marco de...

Risk stratification is particularly important for the prognosis of coronavirus disease 2019 (COVID-19), particularly regarding myocardial injury (MI). Non-contrast chest computed tomography (CT) is useful for diagnosing lung injury, but its potential for one-stop evaluation of MI remains unclear. This study aimed to develop a deep learning (DL) framework for the one-step prediction of MI in COVID-19 patients using non-contrast chest CT. This approach seeks to enable early MI screening, direct high-risk patients toward further diagnostic tests, and optimize the allocation of medical resources.

A group of 453 patients with COVID-19, including 230 patients with MI and 223 patients without MI, were retrospectively recruited. A two-stage DL framework was developed to first segment the left ventricle (LV) in the non-contrast chest CT images using fully convolutional networks with a ResNet-101 backbone (FCN-ResNet-101) module and then classify the status of MI using densely connected convolutional network with the structure type of 121 (DenseNet-121) module. The framework was trained in a training-validation dataset of 413 patients (MI or non-MI) with a cross-validation approach, and evaluated in a testing dataset of 40 patients.

The proposed DL framework accurately obtained the segmentations of LV in non-contrast chest CT images with an intersection over union (IoU) of 0.8041, an accuracy (ACC) of 0.9949, and a Dice coefficient of 0.8672. Based on the segmentation, the DL framework further accurately determined MI status and obtained an area under the curve (AUC) of 0.8618 [95% confidence interval (CI): 0.8049-0.9187], an ACC of 0.7763, a sensitivity (SEN) of 0.8750, a specificity (SPE) of 0.6071, and an F1 score of 0.8317.

DL could determine the status of MI in non-contrast chest CT images of patients with COVID-19, providing one-stop convenience for early screening of MI.