PEFormer: CNN-Transformer Network with Noisy Training for Robust Pulmonary Embolism Detection in Computed Tomography
Pulmonary embolism (PE) is considered the third cause of cardiovascular deaths globally, after stroke and heart attack. It refers to the physical obstruction of pulmonary (lung) arteries by a mass, most commonly a blood clot, which often travels from other parts of the body to the lungs. Due to the symptoms of PE being generic, diagnosis most commonly involves imaging of the chest. Diagnosing PE from Computed Tomography (CT) scans is time-consuming, and research indicates that clinicians over-diagnose PE in patients with chest-related problems. In this study, we propose PEFormer, a method to automatically classify PE from CT scans. The method is composed of a convolutional neural network for slice-level (2D) PE classification, followed by a transformer network with locally windowed self attention for study-level (3D) classification. Furthermore, we propose a semi-supervised noisy student training strategy to incorporate unlabeled images into our training pipeline and improve the ability to generalize well across different datasets. We further improve our proposed Transformer model by adopting a long document classification method from natural language processing (NLP). We utilize the large public RSNA STR Pulmonary Embolism CT Dataset for training and testing and report our results on external opensource datasets. While previous studies achieved promising results, the majority of these works were evaluated on small and private datasets, thereby limiting their generalizability.
A.S.M. Al Mahrooqi, "PEFormer: CNN-Transformer Network with Noisy Training for Robust Pulmonary Embolism Detection in Computed Tomography", M.S. Thesis, Machine Learning, Abu Dhabi, UAE, 2022.