What question did this study set out to answer?

To evaluate how well deep learning models can segment pulmonary embolism in dual-energy CT images and to analyze the relationship between segmentation performance and image quality.

February 26, 2026Open Access

Deep learning models for pulmonary embolism segmentation on dual-energy CT: performance analysis and image quality correlation

Key Points

To evaluate how well deep learning models can segment pulmonary embolism in dual-energy CT images and to analyze the relationship between segmentation performance and image quality.
Conducted a retrospective study on 57 patients' dual-energy CT pulmonary angiography data.
Reconstructed virtual monoenergetic images at 40 keV, 60 keV, 80 keV, and 100 keV, along with iodine maps.
Manually segmented images by radiologists, split into training and test sets after data augmentation.
Applied five deep learning models to segment the data and evaluated using performance metrics such as DSC, IoU, recall, and precision.
Analyzed correlations between image quality metrics and segmentation performance.
In Dataset 1, 80 keV images had the best performance with a DSC of 0.565 ± 0.032.
In Dataset 2, 60 keV images performed best with a DSC of 0.656 ± 0.031.
Attention UNet achieved the highest overall performance among models in both datasets.
Segmentation performance showed strong correlation with subjective opinion score but weak correlation with objective image quality metrics.

Abstract

To evaluate the segmentation performance of dual-energy computed tomography (CT) reconstructed images for pulmonary embolism (PE) and investigate the relationship between segmentation performance and image quality. A retrospective study analyzed dual-energy computed tomography pulmonary angiography (CTPA) data from 57 patients, divided into two datasets (Dataset1 and Dataset2) based on case complexity. Virtual monoenergetic images (VMIs) at 40 keV, 60 keV, 80 keV, and 100 keV, along with iodine map, were reconstructed and manually segmented by radiologists to delineate PE. After augmentation, 18,490 images were split into training and test sets. Five deep learning models were applied to segment different dual-energy reconstructed images, were evaluated using Dice Similarity Coefficient (DSC), Intersection over Union (IoU), Recall, and Precision. Image quality was evaluated, and its correlations with segmentation performance were analyzed using Pearson, Spearman, and Kendall correlation coefficients. In Dataset 1, the 80 keV images achieved the best segmentation performance, with a DSC of 0.565 ± 0.032, recall of 0.596 ± 0.060, IoU of 0.439 ± 0.035, and precision of 0.634 ± 0.059. In Dataset 2, the 60 keV images performed best, with corresponding values of 0.656 ± 0.031, 0.663 ± 0.035, 0.533 ± 0.034, and 0.719 ± 0.034. Among all five segmentation models, Attention UNet consistently achieved the highest performance, with a DSC of 0.545 ± 0.046 on Dataset 1 and 0.649 ± 0.045 on Dataset 2. Segmentation performance was strongly correlated with subjective opinion score, with Pearson, Spearman, and Kendall correlation coefficients of 0.983, 0.800, and 0.667. The 60 keV and 80 keV VMIs consistently demonstrated the most favorable and stable performance trends across datasets of varying complexities, with Attention UNet outperforming the other models. Furthermore, segmentation performance was strongly correlated with subjective opinion score, whereas its correlation with objective image quality metrics such as SNR and CNR was weak and negative.

KI fragen

Bookmark

View Full Paper