This study evaluated deep learning models to differentiate incomplete Kawasaki disease from pneumonia on echocardiographic images. Investigators collected 2D echocardiographic images of the coronary artery in short-axis view from 203 children under six years of age presenting with a high-grade fever, 138 of which had incomplete Kawasaki disease and 65 of which had pneumonia (none had both). Two models were developed, Multiple Receptive Attention Network (MRANet) and Multiple Large Receptive Attention Network (MLRANet), each using receptive attention layers, increased parallel convolutional layers of different dilation rates, and feature maps to detect coronary artery features at multiple scales. 10-fold cross-validation was used to evaluate performance compared to VGG, Xception ResNet, ResNext, SEResNet, SEResNext, and EfficientNet. Classification was assessed based on accuracy, F1 score, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV). MLRANet achieved the best results, with 84.73% accuracy, 93.48% sensitivity, 66.15% specificity, 85.43% PPV, and 82.69% NPV. It also had the highest areas under the curve for precision-recall and receiver operating characteristic curves (0.893 and 0.818 respectively). MRANet was second best and the remaining deep learning algorithms performed at least 10 to 20 points lower in nearly all areas. Using wider and more diverse receptive fields (increasing the number of parallel convolution layers and dilation rates) also enhanced performance. These results were near the accuracy of experienced Kawasaki Disease specialists who have achieved about 85% sensitivity and 70% specificity. Class activation mapping showed that similar to expert cardiologists, MRANet and MLRANet exhibited a higher focus on the coronary arteries within the images in contrast to comparator models. The study was limited by a relatively small data set and the relatively larger incomplete Kawasaki Disease dataset relative to pneumonia.