تشخيص خودكار آمبولي ريوي در تصاوير سي‌تي آنژيوگرافي ريوي

Pulmonary Embolism (PE) is a life-threatening emergency associated with high mortality and morbidity rates. CT Pulmonary Angiography (CTPA) is the diagnostic standard for confirming PE. Early diagnosis and treatment of PE are crucial to preventing complications and reducing mortality. In this study, a novel algorithm comprising a 57-layer convolutional neural network (CNN) based on U-Net, combined with Long Short-Term Memory (LSTM) and Self-Attention (SA), was developed to leverage long-term dependencies in serial data, focus the model on critical information, and reduce unnecessary noise. The dataset was collected from two hospitals in Shiraz (Namazi and Shahid Faghihi) using two different CT scanners, including images from 121 patients (64 females and 57 males). Low-resolution images were retained to enhance generalizability, and data collection was performed randomly. Image preprocessing steps included removing metal artifacts and down-weighting irrelevant areas (e.g., bones) to enhance the model’s focus on emboli. The images were then cropped to extract lung tissue and resized to 256x256 pixels for input into the network. The performance of different eva‎luation metrics was statistically analyzed using SPSS software, with a significance level of P-Value = 0.05. The results demonstrate that the proposed method shows high sensitivity in detecting positive cases and is competitive with similar studies in terms of accuracy and Area Under the Curve (AUC). The slice-based eva‎luation achieved the highest sensitivity of 97.32% [95% CI: 96.34-98.30] and a specificity of 68.35% [95% CI: 64.17-72.54]. On a patient-based scale, the model achieved an average sensitivity of 99.18% and a specificity of 62.4%. The U-Net architecture, due to its symmetric and straightforward structure, has made the model faster and easier to implement, which is a significant practical advantage.

5

حسين عربي