Pulmonary embolism is a life-threatening disease, early detection and treatment can significantly reduce mortality. In recent years, many studies have been using deep learning in the diagnosis of pulmonary embolism with contrast medium computed tomography pulmonary angiography, but the contrast medium is likely to cause acute kidney injury in patients with pulmonary embolism and chronic kidney disease, and the contrast medium takes time to work, patients with acute pulmonary embolism may miss the golden treatment time. This study aims to use deep learning techniques to automatically classify pulmonary embolism in CT images without contrast medium by using a 3D convolutional neural network model. The deep learning model used in this study had a significant impact on the pulmonary embolism classification of computed tomography images without contrast with 85% accuracy and 0.84 AUC, which confirms the feasibility of the model in the diagnosis of pulmonary embolism.
The modern lifestyle of individuals often involves prolonged periods of sitting, whether it's for work, commuting, or leisure activities. This sedentary lifestyle is prevalent in students due to educational demands, as well as in the elderly, where reduced mobility can lead to extended periods of inactivity or confinement to bed. These factors, among others such as pregnancy, uncontrolled cardiovascular diseases, and a history of cancer, can contribute to the development of lower limb venous thrombosis. Notably, studies like that of Suadicani (2012) have shown that lower limb venous thrombosis is the most common cause of Pulmonary Embolism (PE) resulting from blood clots.
In recent years, the field of deep learning has seen significant growth, particularly in applications within the domain of medical imaging. Research focused on employing deep learning to aid in diagnosing PE through computed tomography pulmonary angiography (CTPA) has gained prominence. Studies like that of Huang (2020) have sought to replace manual interpretation of CTPA results with deep learning techniques, aiming to alleviate the clinical workload for medical practitioners. Additionally, Lenfant (2020) have improved the quality of medical images from CTPA scans, thus reducing the radiation dose.
When a physician strongly suspects a patient may have PE, CTPA has become the primary diagnostic method (Gal (2004)). It enables immediate identification and prompt treatment of patients based on medical imaging results. Nevertheless, the risk of underdiagnosing PE remains a significant concern (Yavas (2008)), mainly because of the need for physicians to manually interpret CTPA images. Each patient’s CTPA scan typically comprises approximately 100-300 slices, necessitating extensive time for physicians to meticulously examine each pulmonary artery to detect potential embolic symptoms. Lack of experience, physician fatigue (Joshi (2014)), or inadequate physician coverage can lead to prolonged judgment times and diagnostic errors. Consequently, the healthcare system is burdened with increased stress (Kline (1992), Gerald (2005)). Moreover, as the application of CTPA scans grows, ensuring the provision of accurate and timely diagnostic images becomes increasingly challenging for healthcare systems and physicians (Prologo (2004)).
Patients undergoing CTPA scans are at risk of developing acute kidney injury (AKI), a common occurrence in critically ill patients with PE (Chang (2017)). Once AKI occurs, the patient’s risk of death rises significantly. Additionally, CTPA scans require the administration of contrast medium, and this requires a certain period to allow the contrast medium to circulate through the pulmonary system before the actual CT scan (Bae (2010)). Patients with acute PE may lose valuable treatment time while waiting for the contrast medium to take effect. This delay could be critical since patients with acute PE require immediate diagnosis and treatment.
CTPA images not only serve as the most commonly used method for diagnosing PE but also allow physicians to definitively determine the presence of embolic symptoms in patients without the need for further diagnostic confirmation. This makes CTPA images ideal for supervised learning methods in deep learning for image classification (Weiss (2006)). In recent research, deep learning algorithms used in the medical imaging field predominantly employ supervised learning (Islam (2021)), particularly convolutional neural networks (CNNs) (Ker (2017)). To address healthcare system and physicianrelated issues, Yang et al.’s research (Yang (2019)) proposed a two-stage CNN for assessing PE in CTPA scans. The model exhibited a sensitivity of 75.4% when evaluated using CTPA test data from 20 patients. Huang (2020) introduced PENet, a 3D CNN, for detecting PE. When tested on two datasets, PENet achieved AUROCs of 0.84 and 0.85. These methods represent new avenues for human-robot collaboration (HRC) in healthcare (Kong (2018)). They reduce the probability of diagnostic errors while also decreasing the time physicians spend interpreting images, enabling more appropriate treatment for patients.
While recent studies have used CNNs to diagnose PE in CTPA. Although the accuracy could be improved from a physician’s point of view, these images have been acquired after the administration of contrast medium. CTPA images with contrast medium tend to provide clearer distinctions for diagnosis, as shown in Figure 1. However, contrast medium can have adverse effects on the body. They exert cytotoxic effects on the renal tubular epithelial cells (RTE) of the kidneys, leading to loss of function, severe cell damage, and Figure 1: Illustrates the difference between CTPA scans with and without contrast medium: the left image is without a contrast medium, while the right image is with a contrast medium. In the right image, the circled area clearly shows a darker region. This is because PE causes blockages in the pu
This content is AI-processed based on open access ArXiv data.
