Diagnosis of Pulmonary Embolism from CT Pulmonary Angiogram by Using Deep Transfer Learning and Sobel Filter

Pulmonary embolism (PE) indicates a blood thrombus that travels toward the lungs and can lead to significant morbidity and mortality. Thus, early diagnosis and treatments are crucial. An efficient PE detection mechanism can improve radiologists’ productivity by helping to prioritize critical cases and speed up at-risk patient diagnoses. The standard mechanism for diagnosing PE is the chest computed tomography pulmonary angiogram (CTPA) images. A subset of 15000 samples, including normal and PE-positive lung CTPA images were taken from the largest multinational publicly available RSNA-PE-CT dataset for our purpose. Noise and spatial aliasing in CTPA images can make it challenging to discern boundaries between different anatomical structures or abnormalities; therefore, an edge detection filter (Sobel) was deployed to remove noise and highlight the edges in the original image. In this study, PE in lung CTPA image is detected using deep transfer learning models, which utilize Convolutional Neural Network (CNN), a widely used deep learning (DL) technique. All selected models were trained and validated by adjusting the hyperparameters, which included the layers, loss function, activation function, rate of learning, and epoch number. Nonetheless, compared to other CNN-based frameworks such as MobileNetV2, Xception, InceptionV3, and ResNet50, the customized Sobel Filter-based DenseNet169 (SF-DenseNet169) model performs better. The SF-DenseNet169 model obtained the maximum training accuracy (97.32%) and validation accuracy (96.67%) with minimal loss. As a result, we propose the improved SF-DenseNet169 model to be the effective and reliable means to diagnose PE-positive patients automatically.