Pericardial Effusion Detection on Post-Mortem Computed Tomography (PMCT) Images using Convolutional Neural Network-based Models: Algorithm Development and Validation (Preprint)

BACKGROUND Pericardial effusion and hemopericardium are two common causes of death in the medicolegal setting. Post-mortem computed tomography (PMCT) has become an efficient tool to assist medicolegal death investigation by performing a quick whole-body examination which may avoid the need for traditional autopsy. However, forensic pathologists face challenges in processing a vast amount of PMCT images in practice due to the scarcity of human resources. OBJECTIVE In this work, a Pericardial Effusion Automatic Detection (PEAD) framework is proposed to automatically process many whole-body PMCT images to filter out the irrelevant images and focus on cardiac images, in order to perform pericardial effusion detection. METHODS In the PEAD framework, the standard convolutional neural network architectures of VGG and ResNet are carefully modified to fit the specific characteristics of PMCT images in this work. Comparing the experimental results proves the effectiveness and efficiency of the proposed framework and modified models. RESULTS 2,437 PMCT images were used in the heart region classification experiment while 1,124 PMCT images were used in the pericardial effusion classification experiment. Among the six examined models, the modified VGG8 model has the highest performance in both two tasks. It achieved precision 94.90%, recall 98.50% and 96.33% on testing set for heart region classification. For the pericardial effusion classification, it also achieved precision 94.09%, recall 88.37% and 91.16% on testing set which also overperformed the other five models. CONCLUSIONS In this work, we proposed a framework called PEADS which consist of two main stages. It first filters out the PMCT images that contain the various cardiac structures, and then further classifies whether there is pericardial effusion in the images. Based on the testing set from NMDID, the modified VGG8 shows the most promising performance which suggests it is the best model to utilize and the backbone in the PEAD framework.