A Comparative Study of Deep Learning Models for Detecting Pulmonary Embolism

Abstract

Pulmonary Embolism (PE) is a life-threatening disease caused by the development of a clot in one of the lung arteries. While it is a common disease, it is still hard to diagnose and providing an early diagnosis can improve the odds of survival of the patient drastically. In this paper, we aim to conduct a comparative analysis between a selection of deep learning algorithms that have been trained to detect Pulmonary Embolism(PE) and other exam level indicators which are essential for determining the severity of the disease in a patient. We consider deep learning architectures such as the MobileNet, VGG, Resnet, Xception, Inception, Unet which are extensively used for computer vision tasks and compare their performance based on evaluation metrics such as the loss, accuracy and AUC score. The results obtained during the study shows that streamlined architectures such as the MobileNet, VGG, ResNet and Unet achieve AUC scores of 0.85, 0.85, 0.82 and 0.83 respectively as compared to other models included in the study such as the inception, DenseNet and Xception which achieved an AUC score of 0.5,0.5,0.64 respectively. The links between the increasing and decreasing feature layers of the UNet provide for robust detection of features from an image. Similarly, the streamline depth-wise separable convolution layers architecture present in both the MobileNetV2 and VGG explain the feature detection in the given task. The outcomes of this research also show that there exists a significant difference of confidence between the image level and exam level features. To further support our study, we compare the performance of the selected models with models specifically proposed to detect pulmonary embolisms such as PENet [1] and Pi-PE [2]. Inferences from this methodology suggest that MobileNetV2 and VGG have similar performance as compared to PENet [1], However, Pi-PE [2] outperforms all existing architectures by achieving an AUC Score of 0.91. © 2022, Springer Nature Switzerland AG.