Table 1.
The information of the dataset used in the experiment.
Fig 1.
Schematic illustration of the learning process.
Fig 2.
Accuracy and loss graphs after training the architectures to detect feline HCM using VD X-ray images.
Five DL architectures were used: ResNet50V2, ResNet152, InceptionResNetV2, MobileNetV2, and Xception. Epochs refer to the repetition of learning.
Fig 3.
Confusion matrix from the deep learning process of detecting feline HCM using VD X-ray images.
Five DL architectures were used, with ResNet50V2, ResNet152, and InceptionResNetV2 presented. 22 X-ray images were tested. X-ray images on the right indicate original images that the architecture could not classify.
Fig 4.
Confusion matrix obtained from the deep learning process of detecting feline HCM using VD X-ray images (ResNet50V2, ResNet152, and InceptionResNetV2).
Cont’d. Twenty-two X-ray images were tested. X-ray images on the right indicate original images that the architecture could not classify.
Fig 5.
ROC curve of DL results in detecting feline HCM using VD X-ray images.
Five DL architectures were used and compared (ResNet50V2, ResNet152, InceptionResNetV2, MobileNetV2, and Xception). The AUC was calculated and presented.
Table 2.
Evaluating parameters of each deep neural network model (%)—accuracy, precision, recall, F1 score, AUC score, sensitivity, and specificity.
Fig 6.
Comparison of accuracy of five DL models in old test data, new data, and combined results.
Five DL architectures were used and compared (ResNet50V2, ResNet152, InceptionResNetV2, MobileNetV2, and Xception).
Fig 7.
t-SNE plotting of five deep learning models in all test data.
Fig 8.
Comparison of the accuracy of Majority voting strategy and Softmax strategy in old data, new data, and combined results
(A) and Confusion matrix of Softmax voting strategy and misdiagnosed images (B).