Convolutional architectures with LSTM and TCN to embolism classification: exploring dependency between data

Abstract

Pulmonary Embolism is an affection caused by obstruction of the pulmonary artery or one of its branches. This condition imposes a high mortality incidence, in the United States approximately 100.000 deaths per year. Computed Tomography Pulmonary Angiography is a radiologic modality and an essential technology for diagnosing this disease, providing a series of axial images. We trained two Convolutional Neural Networks (Efficient Net B0 and Resnet 3D 18) in the RSNA-STR Computed Tomography Pulmonary Angiography Dataset to identify this affection. After training these Convolutional Neural Networks, we added a new layer to the architecture by exploring the dependency between the images along the exam using Long Short-Term Memory or Temporal Convolutional Networks. With the models trained and tested, we compared these different approaches using different metrics. As a result, the Temporal Convolutional Network approach with Resnet 3D 18 improved significantly compared to the results found in the other methods. The main contribution of this work was to observe how different combinations of architectures can help classify Computed Tomography Pulmonary Angiography.