Deep Learning for Automatic Pulmonary Embolism Identification Using CTA Images
Hajer Khachnaoui, Mariem Agrebi, Siwar Halouani, Nawrès Khlifa
Abstract
Background; The diagnosis of diseases with Deep Learning (DL) methods such as Convolution Neural Network (CNN) receives spectacular attention due to their efficacy in the learning and classification of complex features. Indeed, DL has been widely used for Pulmonary Embolism (PE) diagnosis. This latter is a serious pathology caused by the presence of a blood clot in the pulmonary arteries. It shows significant mortality and is considered an emergency that requires rapid detection and decision-making. The imaging modality of choice for the detection of this disease is Computed Tomography Angiography (CTA).Method; In this study, an effort has been put to automatically identify PE in CTA images. Firstly, a denoising step was applied to enhance the image quality. Then, the pre-trained VGG-19 and Inception-V3 CNN models are used to classify PE and non-PE images. Next, lungs and lung vessels are segmented from filtred CTA images using mathematical morphology. Finally, a 3D reconstruction of the lungs and lung vessel tree was performed to detect the location of the PE.Results; The proposed method is evaluated on CTA images of 35 patients obtained from a public dataset and divided into training (85%) and validation (15%) sets. We found that Inception-V3 exceeded VGG-19, and reached an accuracy of 90.50%.Conclusions; The proposed method findings represent a successful application of CNN models for the complex task of PE diagnosis in CTA images. It can potentially help radiologists accelerate the diagnostic process and improve care pathways through more efficient diagnosis.