Deep Learning-Based Interpretable Computer-Aided Diagnosis of Drowning for Forensic Radiology

The importance of post-mortem computed tomography (PMCT) is growing as experience and expertise in autopsy develops, with clear evidence to support the use of PMCT in suspected cases of drowning. Our previous work had proposed a computer-aided diagnosis (CAD) system to classify post-mortem lung computed tomography (CT) images into two categories of drowning and non-drowning cases. In this study, we further improved the CAD system in terms of performance and interpretability based on VGG16 deep convolutional neural network (DCNN) via transfer learning. Meanwhile, to provide forensic specialists with interpretable results, visualization was applied to post-mortem lung CT images by means of Gradient-weighted Class Activation Mapping (Grad-CAM) to highlight important regions in the image for predicting the concept. The training and test dataset consist of 293 cases of post-mortem lung CT including 8127 slices. The area under the receiver operating characteristic curve (AUC) reached 0.983 after applying majority vote to slice-based method and the visualization showed that the proposed CAD system was interpretable and feasible for PMCT diagnosis of drowning.