A deep-learning-aided diagnosis of drowning using post-mortem lung computed tomography

• A deep-learning diagnosis for drowning was devised using postmortem CT images. • VGG16 outperformed, achieving the highest mean AUC-ROC of 88.42 %. • Each subject was diagnosed by averaging prediction values from multiple lung slices. • Diagnosis accuracy of 96 % and 79 % were achieved for original and public datasets. Identifying the cause of death using postmortem CT images is crucial since it provides a non-invasive, objective approach for forensic investigations while offering significant advantages in terms of time efficiency and cost-effectiveness compared to traditional autopsy methods. However, due to varied lung conditions in the postmortem CT images, a standardized method to diagnose drowning using CT images has not been established. This study aimed to devise a deep-learning-aided framework for diagnosing drowning from postmortem lung CT images. First, to find the suitable convolutional neural network (CNN) architecture for classifying lung CT images into drowning and non-drowning cases, three well-known CNNs, AlexNet, VGG16, and MobileNet, were trained with a single-institute postmortem CT image dataset and the performance and generalizability were also evaluated using images extracted from a public decedent CT image database. The results showed that VGG16 architecture outperformed the three models with the highest mean AUC-ROC and accuracy values of 88.42 % and 80.56 % respectively for drowning image classification, as well as the highest generalizability with an AUC-ROC of 71.79 % on a public image dataset. Additionally, the case-based diagnosis was performed using probability scores given from the model to each slice taken in the same subject. The final diagnosis accuracy was 96 % on the original dataset and 79 % on the public dataset, showing the strong potential that the devised framework can be used as a screening tool to identify drowning cases using postmortem CT images.