Explainable Artificial Intelligence (XAI): Classification of Medical Thermal Images of Neonates Using Class Activation Maps

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.

The past, current, and future of neonatal intensive care units with artificial intelligence: a systematic review

“… Study Approach Purpose Dataset Type of data (image/non-image) Performance Pros(+) Cons(-) Hauptmann et al, 2019 187 3D (2D plus time) CNN architecture Ability of CNNs to reconstruct highly accelerated radial real‐time data in patients with congenital heart disease 250 CHD patients. Cardiovascular MRI with cine images +Potential use of a CNN for reconstruction real time radial data Lei et al, 2022 158 MobileNet-V2 CNN Detect PDA with AI 300 patients 461 echocardiograms Echocardiography 88% (AUC) +Diagnosis of PDA with AI - Does not detect the position of PDA Ornek et al, 2021 189 VGG16 (CNN) To focus on dedicated regions to monitor the neonates and decides the health status of the neonates (healthy/unhealthy) 38 neonates 3800 Neonatal thermograms 95% (accuracy) +Known with this study how VGG16 decides on neonatal thermograms -Without clinical explanation Ervural et al, 2021 190 Data Augmentation and CNN Detect health status of neonates 44 neonates 880 images Neonatal thermograms 62.2% to 94.5% (accuracy) +Significant results with data augmentation -Less clinically applicable -Small dataset Ervural et al, 2021 191 Deep siamese neural network(D-SNN) Prediagnosis to experts in disease detection in neonates 67 neonates, 1340 images Neonatal thermograms 99.4% (infection diseases accuracy in 96.4% (oesophageal atresia accuracy), 97.4% (in intestinal atresia-accuracy, 94.02% (necrotising enterocolitis accuracy) +D-SNN is effective in the classification of neonatal diseases with limited data -Small sample size Ceschin et al, 2018 188 3D CNNs Automated classification of ...…”

“…DL algorithms have been used to classify neonatal diseases from thermal images 189 – 192 . These studies analyzed neonatal thermograms to determine the health status of infants and achieved good AUC scores 189 – 192 . However, these studies didn’t include any clinical information (Table 6 ).…”

Improving Explainability in CNN-Based Classification of Mask Images with HayCAM+: An Enhanced Visual Explanation Technique

“…Class Activation Mapping (CAM) [30] is a method that uncovers the contributions of pixels using the last convolutional layer. However, CAM requires removing the fully-connected layer and adding Global Average Pooling, which can reduce performance [31]. Gradient-based CAMs (GradCAMs) do not have these limitations and can be applied without modifying the model.…”

HayCAM: A Novel Visual Explanation for Deep Convolutional Neural Networks

“…By freezing the first layers, the model is retrained with the dataset. For the mask classification 2 neurons are added to the pre-trained model as a last layer [49].…”