No abstract

In order to determine the health status of the neonates, studies focus on either statistical behavior of the thermograms' temperature distributions, or just correct classifications of the thermograms. However, there exists always a lack of explain-ability for classification processes. Especially in the medical studies, doctors need explanations to assess the possible results of the decisions. Presenting our new study, how Convolutional Neural Networks (CNNs) decide the health status of neonates has been shown for the first time by using Class Activation Maps (CAMs). VGG16 which is one of the pre-trained models has been selected as a CNN model and the last layers of the VGG16 have been tuned according to CAMs. When the model was trained for 50 epochs, train-validation accuracies reached over 95% and test sensitivity-specificity were obtained as 80.701%-96.842% respectively. According to our findings, the CNN learns the temperature distribution of the body by mainly looking at the neck, armpit, and abdomen regions. The focused regions of the healthy babies are armpit and abdomen whereas of the unhealthy babies are neck and abdomen regions. Thus, we can say that the CNN focuses on dedicated regions to monitor the neonates and decides the health status of the neonates.

No abstract

Monitoring temperature changes of infants in the neonatal intensive care unit is very important. Especially for premature and very low birthweight infants, determining temperature changes in their skin immediately is extremely significant for follow-up processes. The development of medical infrared thermal imaging technologies provides accurate and contact-free measurement of body temperature. This method is used to detect thermal radiation emitted from the body to obtain skin temperature distributions. The purpose of this study is to develop an analysis system based on infrared thermal imaging to classify neonates who are healthy and suffering from heart disease using their skin temperature distribution. In this study, 258 infrared thermograms obtained applying data augmentation on 43 infrared thermograms captured from the Neonatal Intensive Care Unit were used. The following operations were performed: firstly, images were segmented to eliminate unnecessary details on the thermogram. Secondly, the features of the image were extracted applying Discrete Wavelet Transform (DWT), Ridgelet Transform (RT), Curvelet Transform (CuT), and Contourlet Transform (CoT) which are multiresolution analysis methods. Finally, these features are classified as healthy and unhealthy using classification methods such as Artificial Neural Network (ANN), Support Vector Machine (SVM) and Random Forest (RF). The best results were obtained with SVM as 96.12% of an accuracy, 94.05% of a sensitivity and 98.28% of a specificity.

Explaining the decision mechanism of Deep Convolutional Neural Networks (CNNs) is a new and challenging area because of the “Black Box” nature of CNN's. Class Activation Mapping (CAM) as a visual explainable method is used to highlight important regions of input images by using classification gradients. The lack of the current methods is to use all of the filters in the last convolutional layer which causes scattered and unfocused activation mapping. HayCAM as a novel visualization method provides better activation mapping and therefore better localization by using dimension reduction. It has been shown with mask detection use case that input images are fed into the CNN model and bounding boxes are drawn over the generated activation maps (i.e. weakly-supervised object detection) by three different CAM methods. IoU values are obtained as 0.1922 for GradCAM, 0.2472 for GradCAM++, 0.3386 for EigenCAM, and 0.3487 for the proposed HayCAM. The results show that HayCAM achieves the best activation mapping with dimension reduction.