تشخيص و طبقه‌بندي آمبولي ريه در تصاوير سي‌تي آنژيوگرافي ريه با استفاده از روش يادگيري عميق

Pulmonary Embolism Detection an‎d Classification in Computed Tomography Pulmonary Angiography Images Using a Deep Learning Method

مهندسي پزشكي

Pulmonary embolism (PE) is a life-threatening condition requiring rapid an‎d accurate diagnosis. This study proposes a deep learning-based approach for automated PE segmentation, focusing on both pixel-level accuracy an‎d clinical applicability. The aim of this study was to develop a deep learning–based framework for pulmonary embolism segmentation in computed tomography pulmonary angiography (CTPA) images to assist in pulmonary embolism detection an‎d disease severity assessment. The objective is to enhance segmentation performance by incorporating Squeeze-an‎d-Attention (SA) modules within the network architecture an‎d to eva‎luate its generalizability on multi-center datasets. A neural network architecture combining convolutional layers, Long Short-Term Memory (LSTM) units, an‎d SA blocks was developed. The dataset consisted of a public collection (a combination of two online datasets) comprising 126 images, as well as a private dataset including CTPA scans from 34 patients collected from two hospitals in Shiraz, Iran, with variability in imaging devices an‎d patient conditions. Reference masks of pulmonary emboli were generated by two radiologists — one with 12 years of experience an‎d one radiology resident. To augment the dataset, traditional techniques such as rotation, brightness adjustment, an‎d contrast variation were employed. Model eva‎luation was conducted comprehensively at three levels: pixel-level, slice-level, an‎d embolism-level, using both patient-based an‎d slice-based approaches. Performance was assessed using five-fold cross-validation, hold-out testing, an‎d ablation studies. The eva‎luation metrics was assessed using multiple metrics, including Dice Similarity Coefficient (DSC), Jaccard Index, Hausdorff Distance, sensitivity, specificity, area under the ROC curve (AUC), false positive rate (FPR), an‎d false negative rate (FNR). The results demonstrated a Dice similarity coefficient of 69.87% at the pixel level an‎d 98.79% sensitivity at the lesion level on the public dataset, indicating competitive segmentation accuracy. Incorporating SA blocks significantly improved performance, increasing the Dice score by 7.58% an‎d reducing the false positive an‎d false negative rates. Therefore, the proposed framework can be considered a clinically oriented segmentation system that also incorporates pulmonary embolism detection within the lesion analysis process. The model showed strong generalizability across different imaging centers an‎d scanner types. The proposed model demonstrates high accuracy in PE segmentation an‎d robustness across diverse datasets. Its design an‎d eva‎luation framework support its potential utility in clinical settings. Future work will focus on real-time deployment an‎d integration into diagnostic workflows.

5

دكتر حسين عربي

159280