Enhancing diagnostic safety with low iodine, low radiation CTPA classification using deep learning

Abstract

Pulmonary embolism (PE) is a life-threatening condition for which computed tomography pulmonary angiography (CTPA) is the standard diagnostic modality. However, conventional CTPA protocols require relatively high iodine contrast and radiation doses, raising concerns about renal injury and radiation exposure. In this study, we propose a deep learning-based framework for PE diagnosis under low-iodine and low-radiation CTPA conditions. The proposed two-stage framework integrates image enhancement and classification by jointly leveraging original low-exposure images and their super-resolved counterparts. We further construct and publicly release a low-iodine, low-radiation CTPA dataset developed in collaboration with a clinical institution to support reproducible research in safe imaging. Experimental results demonstrate that the proposed method substantially improves diagnostic performance compared with single-branch baselines, achieving an area under the ROC curve (AUC) of 0.928 while maintaining balanced sensitivity and specificity. These findings suggest that the proposed framework enables accurate and safer PE diagnosis under reduced contrast and radiation exposure, offering a practical solution for improving diagnostic safety in clinical CTPA imaging.

Introduction

Pulmonary embolism (PE)^1–3 is a life-threatening cardiopulmonary disorder caused by thrombotic obstruction of the pulmonary arterial system, leading to impaired pulmonary circulation and respiratory dysfunction⁴. Owing to its nonspecific clinical manifestations, PE is frequently underdiagnosed or diagnosed with delay⁵. With increasing clinical awareness and improved imaging availability, PE is now recognized as a prevalent cardiovascular disease with substantial morbidity and mortality^6,7. Early and accurate identification of clinically relevant PE, particularly among intermediate-risk patients, is critical for timely treatment and outcome improvement.

CT pulmonary angiography (CTPA) has become the clinical reference standard for PE diagnosis due to its high sensitivity and specificity^8–10. By directly visualizing intravascular thrombi and pulmonary vascular anatomy, CTPA provides essential morphological information for diagnostic confirmation and risk assessment¹¹. However, conventional CTPA protocols rely on relatively high doses of iodinated contrast agents and radiation exposure to ensure sufficient vascular enhancement, raising concerns regarding contrast-induced nephropathy, allergic reactions, and cumulative radiation risk–particularly in vulnerable patient populations^12,13.

Reducing contrast agent dosage and radiation exposure is therefore an important goal in modern PE imaging. While, low-iodine and low-radiation CTPA acquisitions inevitably suffer from reduced vascular attenuation, increased noise, and suppressed fine-grained structural details. These degradations significantly complicate radiological interpretation and pose substantial challenges for automated image analysis. In particular, deep learning–based PE detection and classification models are highly sensitive to contrast loss, as embolic patterns often manifest as subtle intensity and texture variations that become indistinct under low-exposure conditions. Existing methods, which are typically designed and optimized for standard-dose imaging, often exhibit degraded performance when directly applied to low-contrast CTPA data.

Despite recent progress in deep learning for PE detection and medical image enhancement, there remains a notable lack of robust, end-to-end diagnostic frameworks specifically tailored for low-iodine, low-radiation CTPA. Most prior studies focus either on image enhancement or on embolism classification in isolation, without jointly addressing the intertwined challenges of image quality degradation and diagnostic reliability under low-exposure conditions. This methodological gap limits the clinical applicability of current approaches in safety-oriented imaging protocols.

To address these limitations, we propose a two-stage deep learning framework tailored for pulmonary embolism diagnosis using low-iodine, low-radiation CTPA images. Additionally, we establish and publicly release a CTPA dataset acquired under clinically optimized low-contrast protocols, offering a valuable resource for safe imaging research. Our method integrates a Frequency-Aware Super-Resolution Network (FASRN), which restores contrast and structural detail by enhancing frequency-domain features, with a Dual-Branch Classification Network (DBCN) that fuses complementary representations from original and super-resolved images. This design enables accurate embolism classification while substantially reducing contrast agent requirements, aiming to support safer and more efficient PE assessment in clinical practice.

The major contributions of this work are summarized as follows:

1. Low-Iodine/Low-Radiation CTPA Dataset:We construct and publicly release a clinically validated low-iodine, low-radiation CTPA dataset, providing the first dedicated benchmark for studying safe contrast-minimized PE imaging.

2. Frequency-Aware Super-Resolution Network (FASRN): We introduce a novel frequency-enhanced super-resolution model that adaptively restores contrast and structural details in low-exposure CTPA, improving vessel clarity and emboli visibility.

3. Dual-Branch Classification Network (DBCN): We develop a dual-path classification architecture that jointly exploits original and super-resolved images through cross-attention fusion, capturing complementary multi-scale representations for robust PE diagnosis.

4. Clinically Comparable Performance Under Low Contrast: Our integrated framework achieves diagnostic accuracy comparable to standard-dose CTPA while operating on significantly reduced contrast-agent input, highlighting its clinical practicality and safety benefits.

5. Comprehensive Evaluation and Benchmarking: We conduct extensive quantitative and qualitative analyses, including comparisons with state-of-the-art enhancement and classification models, demonstrating the superiority of our approach across multiple evaluation metrics.

The remainder of this paper is organized as follows. Section “Related Work” reviews recent developments in CTPA analysis, image enhancement, and deep learning-based PE detection. Section “Materials” describes the dataset construction, acquisition protocol, and data preparation. Section “Methods” details the proposed FASRN and DBCN architectures. Section “Results and Discussion” presents experimental findings and comparisons with existing approaches. Finally, Section “Conclusion” summarizes our contributions and discusses future directions.

Related work

Recent studies have applied deep learning techniques to automated pulmonary embolism (PE) detection and CT pulmonary angiography (CTPA) analysis. Early approaches mainly utilized conventional convolutional neural networks (CNNs) to extract discriminative features from axial slices or volumetric CT data, achieving promising classification and detection performance under standard-dose, high-contrast imaging conditions^14,15. Subsequent work introduced attention mechanisms and multi-scale feature extraction strategies, enabling improved localization of emboli and suppression of irrelevant structures within complex pulmonary vasculature^16,17. While these methods demonstrate the potential of deep learning to assist radiologists and reduce diagnostic errors, they are generally developed and evaluated under standard acquisition protocols and remain sensitive to contrast degradation and noise, limiting their robustness in low-iodine, low-radiation CTPA settings.

Parallel research has focused on image enhancement and super-resolution techniques to compensate for reduced image quality. Methods such as denoising, contrast adjustment, and frequency-aware reconstruction have been applied to low-dose or low-contrast CT scans, showing enhanced visualization of small anatomical structures and improved downstream classification performance^18–22. However, most existing super-resolution and enhancement approaches are optimized for perceptual quality metrics or visual appearance, rather than diagnostic discriminability. As a result, they may over-smooth subtle embolic features or introduce hallucinated textures that are not clinically meaningful, particularly when applied independently of the downstream diagnostic task.

A further line of research addresses low-dose and low-contrast CT acquisition, aiming to reduce radiation exposure and contrast agent load while maintaining diagnostic accuracy^12,13. These studies highlight that lower contrast dosage significantly reduces vascular visibility, posing challenges for both human interpretation and automated analysis. Although deep learning models have been explored for reconstruction or classification under such conditions, most existing methods treat image enhancement and PE classification as separate problems and adopt single-branch architectures that rely on a single image representation, limiting their ability to balance anatomical fidelity and enhancement-induced artifacts.

In contrast to prior work, the present study targets the unique challenges of low-iodine, low-radiation CTPA by integrating image enhancement and classification within a unified, task-aware framework. By combining frequency-aware super-resolution with a dual-branch classification strategy that jointly exploits original and enhanced images, the proposed method aims to improve diagnostic robustness and reliability under clinically constrained, low-exposure imaging conditions.

Materials

Description of the low-iodine, low-radiation CTPA dataset

All experiments in this study were conducted on the proposed low-iodine, low-radiation CTPA dataset, consisting of real clinical scans collected at Beijing Hospital under reduced contrast agent and radiation protocols. The dataset includes a total of 191 adult patients who underwent CTPA examination for suspected pulmonary embolism, comprising 95 PE-negative cases, 79 cases with acute pulmonary embolism, and 17 cases with chronic pulmonary embolism. Each patient study contains approximately 400 axial CTPA slices. In addition, a subset of 30 PE-positive patients with a total of 773 manually annotated slices was constructed to support fine-grained lesion-level analysis.

All patients were scanned using a 320-row wide-detector CT scanner (Canon, Japan) in the supine position with arms raised. Scans were acquired in helical mode with a tube voltage of 100 kVp and automatic tube current modulation (Sure Exposure 3D) using a preset noise index of 11.5. Additional acquisition parameters included a detector collimation of  mm, pitch of 0.813, gantry rotation time of 0.5 s/r, field of view of 280 mm, and a matrix size of . Images were reconstructed with a slice thickness and interval of 1.0 mm using a standard vascular reconstruction kernel (FC18) and advanced iterative reconstruction algorithms: AIDR3D (Adaptive Iterative Dose Reduction 3D) or FIRST (Forward-projected model-based Iterative Reconstruction SoluTion), which provide dose-dependent noise reduction and improved image quality, particularly for low-exposure acquisitions.

To quantify radiation exposure, the mean tube current was 220.2 ± 84.5 mA for FIRST reconstructions, compared with 264.1 ± 81.2 mA for AIDR3D reconstructions (P = 0.020). The dose-length product (DLP) for FIRST scans was 99.7 ± 28.7 mGycm, lower than the 118.1 ± 29.5 mGycm observed with AIDR3D, and the corresponding estimated effective dose (ED) was 1.39 ± 0.49 mSv versus 1.65 ± 0.41 mSv, representing a 15.7% reduction in ED. Tube current showed a positive correlation with patient BMI in both groups (r = 0.82 for FIRST, r = 0.80 for AIDR3D, P < 0.001). These metrics indicate that the proposed low-exposure protocol achieves clinically meaningful dose reduction while maintaining image quality suitable for PE diagnosis. For contrast enhancement, a dual-head power injector was used to administer iodinated contrast agent (iopamidol, 350 mg/ml) via an antecubital vein at an injection rate of 5.0 ml/s, corresponding to an iodine delivery rate of 1.75 g/s. A reduced contrast volume of 30 ml was employed, corresponding to a total iodine load of 10.5 g, followed by an equal-volume saline flush at the same rate. Bolus tracking was performed with the region of interest placed in the main pulmonary artery at the level of the tracheal bifurcation, and image acquisition was triggered once the attenuation reached 180 HU, with a fixed delay of 4 s. Compared with commonly reported standard CTPA protocols in the literature, which typically use 50–100 ml of contrast agent, the proposed protocol substantially reduces total iodine load while maintaining diagnostic image quality.

Radiation dose metrics, including the dose-length product (DLP), were automatically recorded by the CT system after each scan. The effective radiation dose (ED) was estimated using a standardized conversion coefficient of  mSv/(mGycm) in accordance with European Commission guidelines. The proposed low-exposure protocol achieved a notable reduction in radiation dose compared with conventional CTPA acquisition strategies reported in prior studies.

All CTPA images were independently reviewed by two radiologists, including one deputy chief physician and one resident physician, using RadiAnt DICOM Viewer (version 2024.2) on axial images with a slice thickness of 1.0 mm. Pulmonary embolism presence and extent were assessed according to the Mastora scoring criteria. In cases of disagreement, a consensus reading was performed through joint discussion to determine the final diagnosis. This consensus-based labeling strategy was adopted to ensure reliable ground-truth annotations for subsequent model training and evaluation.

Overall, this dataset provides a clinically realistic and well-characterized benchmark for developing and evaluating deep learning models for pulmonary embolism diagnosis under low-iodine, low-radiation CTPA imaging conditions.

Study design and participants

This retrospective observational study aimed to evaluate the diagnostic performance of a deep learning-based two-stage classification framework for detecting pulmonary embolism (PE) in low-iodine, low-radiation contrast-enhanced chest computed tomography (CT) images. Two complementary datasets were used: real clinical low-exposure CT scans and simulated low-exposure images generated from a public dataset, to comprehensively evaluate model robustness, cross-domain adaptability, and clinical applicability.

Clinical data were obtained from Beijing Hospital and included anonymized contrast-enhanced chest CT scans of adult patients who underwent imaging for suspected PE between February 2022 and July 2023. All patient data were fully anonymized and labeled by experienced radiologists based on confirmed diagnostic outcomes. Inclusion criteria required complete scan protocols and verified diagnoses, while exclusion criteria included severe motion artifacts, missing metadata, non-standard contrast dosing, or non-thoracic scan protocols.

To assess cross-domain generalization, the publicly available RSNA Pulmonary Embolism Detection dataset²³ was also incorporated. All images from this dataset were used in accordance with its licensing terms. To enable comparison under matched imaging conditions, simulated low-iodine and low-radiation images were generated from the full-dose RSNA scans using a controlled contrast attenuation and noise-based simulation strategy, as described in the Data Preprocessing section.

Ethics statement

This retrospective study was approved by the Beijing Hospital Ethics Committee (2024BJYYEC-KY089-02). The requirement for informed consent was waived by the Ethics Committee due to the retrospective design and the use of fully anonymized clinical data. All methods were performed in accordance with the relevant guidelines and regulations, including the Declaration of Helsinki.

Data preprocessing

All clinical CTPA images were reconstructed at a native in-plane resolution of 512512 pixels per slice. To ensure consistency across datasets, both the Beijing Hospital scans and the RSNA Pulmonary Embolism Detection dataset were uniformly preprocessed. This included resampling all images to a consistent voxel spacing, intensity normalization to the [0, 1] range, and spatial alignment to mitigate scanner-dependent variations.

For the RSNA dataset, simulated low-iodine and low-radiation images were generated from the original full-dose CTPA scans by applying a dose-scaling transform to attenuate contrast enhancement, followed by the addition of signal-dependent and Gaussian noise to approximate low-dose acquisition conditions. This simulation strategy reduces vascular iodine signal while maintaining underlying anatomical structures, allowing controlled evaluation of model robustness under degraded imaging conditions.

After basic preprocessing, each slice underwent image-quality enhancement using a frequency-aware super-resolution module to improve visualization of vascular boundaries and subtle thrombus patterns under low-exposure conditions. The enhanced slices were then partitioned into 256256 patches with 50% overlap (stride = 128 pixels) to preserve local structural details while maintaining computational efficiency. Standard data augmentation techniques, including horizontal flipping and small random rotations, were applied during training. Patch-level model outputs were aggregated through confidence-weighted pooling at inference time to obtain image-level predictions.

The proposed low-iodine, low-radiation CTPA classification method

To achieve reliable pulmonary embolism classification under low-iodine, low-radiation CTPA imaging conditions, we design a two-stage deep learning framework that integrates image enhancement and diagnostic feature learning. As shown in Fig. 1, the framework first applies a super-resolution network to improve vessel visibility and restore fine structural details in low-exposure images. The enhanced images are then fed into a dual-branch classification network, which jointly analyzes the original and super-resolved images to capture complementary global and local features. This two-stage design synergistically improves both image quality and classification accuracy, enhancing robustness to low-contrast imaging while maintaining interpretability.

Fig. 1.

Architecture of the proposed two-stage low-iodine, low-radiation CTPA classification framework. The framework consists of a Frequency-Aware Super-Resolution Network (FASRN) for image enhancement and a Dual-Branch Classification Network (DBCN) for pulmonary embolism diagnosis.

In particular, in the enhancement stage, we propose a Frequency-Aware Super-Resolution Network (FASRN), which improves image quality by amplifying diagnostically critical high-frequency components, such as vascular contours and thrombus boundaries, while suppressing low-frequency noise. The network first extracts spatial features from the input CTPA slices using multi-layer convolutional and transformer blocks. These features are then transformed into the frequency domain via the discrete cosine transform (DCT)²⁴, where a frequency-aware attention module selectively enhances important frequency components. Finally, the enhanced image is reconstructed through inverse DCT, combining both spatial and frequency information. By jointly modeling spatial and frequency representations, FASRN effectively enhances fine vascular textures while suppressing low-frequency artifacts, achieving high-fidelity structural restoration under low-exposure conditions.

For robust embolism recognition, we design a Dual-Branch Classification Network (DBCN) that jointly learns from both the original and super-resolved images. The original-image branch focuses on global contextual and structural information, while the super-resolved branch captures fine local details and textures. To enhance discriminative representation, Efficient Channel Attention (ECA)²⁵ modules are embedded in both branches to model inter-channel dependencies, and patch attention modules are introduced to emphasize local discriminative regions. Subsequently, features from both branches are then integrated via cross-attention fusion, allowing complementary information exchange. This design not only leverages the enhanced information but also mitigates the risk of over-dependence on synthetic or over-smoothed features from the enhancement stage. As a result, DBCN can effectively detect small and subtle embolic regions even under low-iodine, low-radiation conditions, achieving stable and interpretable classification.

The entire two-stage framework is trained and validated on both real clinical low-iodine, low-radiation CTPA dataset and simulated public dataset²³. In the enhancement stage, FASRN is trained using real low-exposure images to recover high-frequency details and improve contrast fidelity. In the classification stage, DBCN is first pre-trained on a large-scale normal-dose CTPA dataset to establish initial diagnostic capability and then fine-tuned on real low-exposure and super-resolved image pairs. Binary cross-entropy loss is employed as the supervisory objective, and threshold optimization is conducted based on validation-set F1-score and AUC to maximize classification accuracy.

Implementation details

The proposed framework was implemented in Python (version 3.9) using the PyTorch deep learning library (version 2.6.0). All experiments were conducted on a workstation equipped with an NVIDIA RTX A6000 GPU (48 GB memory, CUDA version 12.4) and an Intel Xeon Gold 6226R CPU, running Ubuntu 20.04 LTS.

FASRN is trained using the Adam optimizer with an initial learning rate of , reduced in a multi-step manner at 200k, 320k, 360k, and 380k iterations. A batch size of 8 is used, and the network is trained for 400k total iterations with an L1 loss to promote high-frequency detail recovery and contrast uniformity. After FASRN training is completed, DBCN is trained using enhanced–original image pairs. The classification model adopts a batch size of 8 patches, patch size of 256256 pixels with 50% overlap, and early stopping based on validation AUC to prevent overfitting. BCE loss is used as the primary objective, with focal loss examined in ablation studies to address class imbalance between positive and negative embolism samples. The Adam optimizer is used for DBCN training with a learning rate of for newly initialized layers and for backbone fine-tuning. The training is conducted for up to 20 epochs, with deterministic behavior ensured by fixing random seeds (42 for Python, NumPy, and PyTorch) and disabling CuDNN benchmarking.

Model training follows a pretraining–fine-tuning workflow: DBCN is first pretrained on the simulated low-exposure RSNA subset to establish baseline diagnostic capability, and then fine-tuned on the real clinical dataset to align with true acquisition characteristics. For internal evaluation, the clinical dataset is partitioned at the patient level into training (70%), validation (15%), and test (15%) sets. During the training of the classification model, positive and negative samples are organized and sampled using different strategies. For positive cases, a total of 773 manually and finely annotated CTPA slices from 30 PE-positive patients are used. These patients are split at the patient level into training and validation sets (70%/15%). This annotated subset is used exclusively for model training and validation and is not included in the final test set to avoid potential information leakage. Negative samples are obtained from 95 PE-negative patients and are partitioned at the patient level into training, validation, and test sets following the same 70%/15%/15% ratio. During the training and validation phases, to balance the number of positive and negative samples and to control computational complexity, a limited number of slices are randomly sampled from each negative patient’s CTPA scan, forming a total of 1,500 negative samples for training and validation. During testing, positive cases are drawn from the remaining PE-positive patients in the full clinical dataset, excluding the 30 patients with manually annotated slices, ensuring complete patient-level independence between the training/validation and test sets. During inference, each CTPA slice is divided into overlapping patches, and patch-level predictions from the dual-branch classifier are aggregated using confidence-weighted pooling to yield a single scan-level (patient-level) PE classification result. All performance metrics reported in this study are computed at the patient level, consistent with clinical diagnostic practice. The exact split lists are available upon request to ensure full reproducibility of the reported results.

Experiments and results

Model performance was evaluated using image-level metrics, including area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, and F1 score. For the clinical dataset, these metrics were computed on the held-out test set following a patient-level split of 70% training, 15% validation, and 15% test. For the RSNA dataset, all evaluation metrics were calculated using the entire test set.

FASRN was assessed using image quality metrics, including peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM), computed between the SR outputs and the corresponding full-resolution reference images. These metrics provide quantitative measures of the SR module’s ability to restore fine anatomical structures while preserving overall image fidelity.

To quantify uncertainty in the classification performance, 95% confidence intervals (CIs) were estimated using bootstrapping with 1000 resamples. For each resample, the metrics were recalculated, and the resulting distributions were used to compute the 95% CIs for AUC, sensitivity, specificity, accuracy, and F1 score.

Model robustness was evaluated by testing performance on noisy versions of the input images. The relative changes in AUC and accuracy compared to the clean images were calculated to quantify the impact of added noise. These results provide insight into the model’s reliability under degraded imaging conditions, such as those encountered in low-exposure CT scans.

Performance evaluation of FASRN

To evaluate the effectiveness of the proposed Frequency-Aware Super-Resolution Network (FASRN), we conducted comprehensive experiments on low-iodine, low-radiation CTPA images. PSNR and SSIM metrics were computed only on the simulated RSNA low-exposure image pairs, where corresponding full-resolution reference images exist, and are reported in Table 1. These metrics do not reflect quantitative fidelity on the real proposed low-iodine, low-radiation CTPA dataset, as truly paired high-resolution references are not available. As shown in Table 1, FASRN substantially improves image reconstruction quality compared with conventional spatial-domain and generic super-resolution approaches. This improvement primarily stems from FASRN’s innovative dual-domain learning strategy, which jointly models spatial and frequency representations through a frequency-aware attention mechanism.

Table 1.

Comparison of different super-resolution methods on the simulated low-exposure CTPA dataset.

Method	PSNR (dB)	SSIM
bicubic26	28.85	0.8997
Lanczos27	29.03	0.8956
LapSRN28	29.75	0.8938
ESPCN29	29.90	0.8457
EDSR30	29.91	0.9000
FSRCNN31	30.01	0.8806
FASRN(ours)	33.29	0.9472

Pulmonary embolism CTPA scans typically exhibit intricate vascular textures, subtle density variations, and indistinct lesion boundaries. Traditional spatial-domain super-resolution approaches, whether convolution-based or Transformer-based, often struggle to recover high-frequency structural details, leading to over-smoothed vessel edges and degraded lesion contrast. In contrast, FASRN integrates multi-layer convolutional operations with Frequency-Transform Blocks (FTBs) to capture both local spatial context and long-range structural dependencies, producing anatomically coherent reconstructions that preserve vascular continuity and realism.

By transforming intermediate feature representations into the frequency domain via the Discrete Cosine Transform (DCT), FASRN explicitly decomposes image information into distinct frequency components. The Frequency-Aware Module (FAM) adaptively amplifies diagnostically critical high-frequency regions, such as vessel contours and thrombus boundaries, while suppressing redundant low-frequency noise. This dual-domain representation learning significantly enhances diagnostic fidelity by preserving fine vascular cues and maintaining global anatomical consistency.

In contrast to generic pre-trained super-resolution models, FASRN undergoes domain-specific fine-tuning on low-exposure CTPA data, enabling it to adapt to the grayscale distribution and texture complexity unique to medical imaging. This domain-aware optimization achieves a robust balance between noise suppression and detail preservation, thereby improving both image realism and interpretability from a diagnostic perspective.

Collectively, FASRN’s dual-domain feature fusion, frequency-adaptive attention, and domain-aware optimization work synergistically to restore vascular continuity and delineate embolic boundaries with high fidelity. As shown in Table 1, FASRN achieves the highest quantitative scores among all compared methods, with a PSNR of 33.29dB and an SSIM of 0.9472. The qualitative visual results in Fig. 2 further confirm that FASRN produces sharper vessel boundaries, clearer thrombus regions, and overall more diagnostically faithful reconstructions than conventional approaches.

Fig. 2.

Visual comparison of super-resolution results.

Classification performance with dual-branch model

To assess the contribution of the proposed FASRN to pulmonary embolism (PE) classification under low-iodine, low-radiation CTPA imaging conditions, we conducted comparative experiments using three input settings: (1) original low-exposuree images (Org), (2) SR-enhanced images, and (3) a dual-branch model that integrates features from both. The goal was to determine whether FASRN could provide complementary information to improve diagnostic accuracy. As shown in Table 2, the baseline model on original low-exposure images achieved relatively high sensitivity (81.11%) but moderate specificity (69.47%), indicating a mild bias toward positive predictions. To assess the statistical significance of the performance improvement, we performed DeLong’s test to compare ROC AUCs between the proposed dual-branch model and the baseline models on the same test set. The results indicate that the AUC improvement achieved by the proposed method is statistically significant (p < 0.001). When using SR-enhanced images alone, although more local details were recovered, subtle structural inconsistencies and potential noise amplification led to a marginal decline in AUC (0.775, 95% CI: 0.737–0.812.737.812) compared with the original images (AUC = 0.814, 95% CI: 0.778–0.850.778.850). In contrast, the proposed dual-branch classification framework effectively combines the global structural stability of the original images with the fine-grained vascular and texture features from SR reconstructions. This hybrid representation significantly improved overall diagnostic performance (AUC = 0.928, 95% CI: 0.904–0.949.904.949), achieving a more balanced trade-off between sensitivity and specificity (F1 = 79.50%). These results indicate that while SR images alone do not surpass the diagnostic capability of original low-exposure CTPA, they provide complementary discriminative cues that enhance embolic lesion detection when jointly modeled. The dual-branch approach thus achieves superior classification robustness and interpretability, demonstrating its potential for safer and more reliable diagnosis under reduced contrast-agent conditions.

Table 2.

Diagnostic performance of baseline and dual-branch models on low-exposure CTPA dataset. Metrics include AUC, sensitivity, and specificity with 95% confidence intervals (CI), as well as precision, recall, and F1-score. Higher values indicate better performance.

Model	AUC (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)	Precision	Recall	F1-score
Baseline (Org)	0.814 (0.778–0.850.778.850)	82.80% (77.00–88.14)	68.52% (63.86–73.33)	57.25%	81.11%	67.13%
Baseline (SR)	0.775 (0.737–0.812.737.812)	74.88% (69.00–81.14)	65.44% (60.64–70.39)	55.35%	66.11%	60.25%
Dual-branch Model	0.928 (0.904–0.949.904.949)	88.34% (83.15–92.57)	82.97% (79.05–86.72)	72.27%	88.33%	79.50%

Recent deep learning models for pulmonary embolism detection, such as hybrid CNN-ViT frameworks^32,33, have achieved high accuracy (up to 97.8%) on standard-dose CTPA datasets with full iodine contrast. These methods are based on sufficient vascular enhancement and relatively low image noise, and their performance under low-iodine, low-radiation imaging conditions has not been systematically studied. In contrast, our dual-branch framework is specifically designed for low-exposure CTPA scans, combining original and super-resolution, enhanced features to compensate for reduced contrast and increased noise. Despite the challenging imaging conditions, it achieves an AUC of 0.928, demonstrating that reliable PE detection is feasible even when standard-dose imaging is not possible. This highlights that our method complements existing state-of-the-art models by extending PE detection capabilities to safer, low-exposure protocols rather than merely optimizing performance under ideal acquisition conditions.

It is worth noting that the SR-only input underperforms the baseline model using original low-exposure images (AUC 0.775 vs. 0.814). This phenomenon can be attributed to subtle structural inconsistencies and potential noise amplification introduced during super-resolution reconstruction. Although SR outputs appear visually sharper, such local distortions may mislead the classification network, particularly in regions with small or low-contrast emboli. The dual-branch model mitigates this risk by integrating complementary information from both the original and SR-enhanced images. Confidence-weighted aggregation and attention-guided feature fusion allow the network to emphasize diagnostically reliable regions while down-weighting areas susceptible to SR-induced artifacts. From a clinical safety perspective, the SR module functions as an auxiliary enhancement rather than a standalone diagnostic input, ensuring that the final patient-level PE decision remains anchored to the original image data. This design helps prevent “hallucination-like” features from adversely affecting high-stakes diagnostic outcomes while still benefiting from the additional fine-grained information provided by super-resolution.

To further validate the discriminative capability of the proposed DBCN, we compared receiver operating characteristic (ROC) and precision-recall (PR) curves, as shown in Fig. 3. The ROC analysis shows that the dual-branch model achieves a significantly higher area under the curve and remains closer to the ideal top-left corner, reflecting stronger global classification capability. In PR analysis, the dual-branch model consistently maintains higher precision across high-recall regions, indicating improved robustness in detecting small or subtle embolic lesions under low-exposure conditions. Collectively, these results demonstrate that SR-based enhancement not only refines image details but also provides complementary diagnostic features, leading to more stable and accurate PE classification with reduced contrast usage.

Fig. 3.

Comparative analysis of model performance using ROC and PR curves. The dual-branch hybrid model demonstrates superior classification ability, with its ROC curve closer to the top-left corner and its PR curve maintaining higher precision in the high-recall region, compared to single-input baselines.

Ablation Study of the Dual-Branch Fusion Mechanism. To better understand why the proposed dual-branch framework outperforms both single-input baselines and to isolate the contribution of the fusion strategy itself, we conducted an ablation study by progressively removing key components of the fusion design. The results are summarized in Table 3. First, we evaluated a naive dual-branch variant in which features extracted from the original and SR-enhanced images were directly concatenated without confidence-weighted or attention-guided fusion. Although this simple concatenation improved performance compared with the SR-only baseline, it remained clearly inferior to the full dual-branch model. This result indicates that performance gains cannot be attributed solely to the availability of additional SR information, but rather depend on how the two representations are integrated. Next, we removed the confidence-weighted gating mechanism while retaining the dual-branch architecture. In this setting, features from the SR branch contributed to the final representation with a fixed weighting. While this variant further improved AUC relative to naive fusion, it still underperformed the complete model. This observation highlights the importance of dynamically regulating the contribution of SR features, especially under low-iodine and low-radiation conditions where SR reconstructions may introduce visually plausible but diagnostically unreliable details. Overall, the ablation results demonstrate that the proposed fusion strategy plays a critical role in balancing enhanced local detail and global anatomical consistency. By adaptively suppressing potentially misleading SR-induced features while preserving stable cues from the original low-exposure images, the full dual-branch model achieves superior diagnostic performance and robustness. These findings support the design choice of treating SR enhancement as an auxiliary source of information rather than a standalone replacement, which is particularly important for maintaining clinical safety in high-stakes PE diagnosis.

Table 3.

Ablation study of fusion strategies in the dual-branch framework.

Model	AUC	Sensitivity	Specificity
Concatenation Model	0.855	68.89%	82.07%
Dual-branch without Confidence Gating	0.899	81.11%	80.67%
Proposed Dual-branch Model	0.928	88.34%	82.97%

Failure case analysis. To further analyze why SR-only inputs underperform the original low-exposure images, we examined representative failure cases in which the baseline model using original images produced correct predictions, while the SR-only model yielded false positives or false negatives. As shown in Figure 4, SR reconstruction enhanced local vessel sharpness and contrast, but also amplified irregular high-frequency patterns along vessel boundaries or adjacent soft tissues. Although visually plausible, these patterns did not correspond to true embolic lesions and occasionally disrupted the learned decision boundary of the classifier. In contrast, the proposed dual-branch model correctly classified a substantial portion of these failure cases. By jointly leveraging the original low-exposure images and the SR-enhanced reconstructions, the model preserved the global anatomical consistency and stable intensity distribution from the original images, while selectively incorporating complementary fine-grained features from the SR branch. This fusion mechanism reduced the influence of spurious SR-induced details and improved robustness against visually convincing but diagnostically misleading features. These observations suggest that SR enhancement alone may introduce subtle distortions that are difficult to distinguish from true pathological signals when used as the sole input. Treating SR images as an auxiliary representation rather than a replacement allows the model to benefit from improved local detail while maintaining diagnostic faithfulness. Clinically, this design is consistent with routine practice, where PE diagnosis relies on multi-slice patient-level assessment and radiologist verification, providing an additional safeguard against potential SR-induced misinterpretation.

Fig. 4.

Failure cases of SR-only classification and recovery by the dual-branch model.

Noise robustness analysis

To further evaluate the stability and reliability of the proposed dual-branch classification framework under realistic imaging conditions, we assessed its robustness to varying levels of noise contamination. Since low-iodine, low-radiation CTPA acquisitions are often accompanied by increased noise, we simulated signal-to-noise ratio (SNR) degradation by 10%, 20%, and 30% relative to the original validation set. Gaussian-Poisson mixed noise was applied to both the original and super-resolved branches without retraining the model, allowing an unbiased evaluation of inference robustness.

As summarized in Table 4, the model maintained stable classification performance across all noise levels. The AUC decreased only marginally from 0.9275 to 0.9132 as SNR dropped from 1.0 to 0.7, while the F1-score remained near 0.80 and sensitivity consistently exceeded 0.86. These results indicate that the proposed framework preserves strong discriminative capability even under substantial image noise.

Table 4.

Noise robustness evaluation of the proposed dual-branch hybrid model under different signal-to-noise ratio (SNR) levels. The model maintains high discriminative performance even with 30% SNR degradation.

SNR	AUC	Sensitivity	Specificity	F1-score
1.0	0.9275	88.33%	82.91%	79.50%
0.9	0.9166	88.33%	85.15%	81.12%
0.8	0.9142	89.44%	82.63%	79.90%
0.7	0.9132	86.11%	85.43%	80.10%

This robustness can be attributed to the confidence-based feature fusion and attention-guided dual-branch design, which jointly suppress background interference and enhance informative vascular cues. By leveraging complementary spatial and frequency representations, the model effectively distinguishes thrombotic structures from noise artifacts, ensuring reliable embolism recognition in low-iodine, low-radiation imaging scenarios.

Discussion

This study proposes a two-stage deep learning framework for pulmonary embolism (PE) detection under low-iodine, low-radiation CTPA imaging conditions. In the first stage, a super-resolution (SR) module enhances the images to recover fine vascular structures. In the second stage, a dual-branch classification model integrates complementary features from both the original and SR-enhanced images, thereby mitigating the diagnostic challenges associated with reduced contrast agent dosage while maintaining clinically acceptable performance.

Our results demonstrate that the framework effectively compensates for image quality degradation in low-iodine, low-radiation CT acquisitions. On a real low-exposure CTPA test set, the dual-branch model achieved an AUC of 0.928, sensitivity of 88.3%, specificity of 82.9%, and an F1 score of 79.5%, outperforming the baseline model that used only the original low-exposure images. Quantitative analyses indicate that the SR module plays a pivotal role in restoring vascular details and thrombus boundaries, while the dual-branch feature fusion further enhances classification robustness by integrating multi-scale information under challenging imaging conditions. Furthermore, the noise robustness experiments confirmed that the model maintains stable classification performance even under degraded SNR levels, highlighting its potential reliability in low-exposure CTPA environments.

Beyond performance improvements, the proposed framework also addresses the clinical challenges associated with low-iodine, low-radiation CTPA protocols. PE diagnosis predominantly relies on contrast-enhanced imaging, yet conventional high-dose contrast and radiation strategies pose risks for patients with renal insufficiency, advanced age, cardiac dysfunction, or critical illness, limiting the applicability of standard CTPA. Although lowering radiation and iodine dosage reduces adverse effects, it also compromises vascular contrast and thrombus visibility. By integrating super-resolution enhancement with dual-branch feature learning, our method effectively alleviates these issues and maintains diagnostic performance comparable to full-dose imaging. On low-exposure scans, the model achieved an AUC of 0.928 (95% CI 90.4–94.9), significantly higher than the baseline AUC of 81.4 (95% CI 77.8–85.0), with sensitivity and specificity approaching those observed under conventional protocols. These findings indicate that the SR module preserves subtle vascular and thrombus structures, while the dual-branch classifier leverages complementary information to sustain diagnostic confidence under reduced exposure conditions.

From a clinical perspective, this approach offers a practical and safer pathway for deploying low-iodine, low-radiation CTPA in routine practice. By algorithmically compensating for quality degradation, it reduces the risk of contrast-induced nephropathy and other adverse reactions in vulnerable patient groups. Furthermore, its modular design is compatible with existing imaging workflows, supporting seamless integration into PACS systems or CT post-processing pipelines. Collectively, these results suggest that deep learning–based enhancement and classification can improve both the safety and accessibility of PE assessment, facilitating personalized, low-risk, and efficient imaging in real-world settings.

However, several limitations remain. First, part of the training data was derived from simulated low-iodine, low-radiation CTPA images generated from normal-dose scans. Although such simulations enable controlled data augmentation and facilitate model pretraining, they cannot fully reproduce the complex characteristics of real clinical low-exposure acquisitions, such as patient-specific contrast dynamics, heterogeneous noise patterns, and scanner-dependent artifacts. This discrepancy may introduce a domain gap and potentially affect model generalization when applied to real-world low-exposure CTPA scans. To mitigate this limitation, the proposed framework does not rely solely on simulated data. Instead, the classification network is subsequently fine-tuned and evaluated on real clinical low-iodine, low-radiation CTPA scans, allowing the model to adapt to authentic imaging characteristics. Nevertheless, simulated data cannot entirely replace real low-exposure acquisitions, and future work will focus on expanding multi-center real-world datasets and exploring domain adaptation strategies to further reduce the gap between simulated and clinical imaging conditions. Second, the current model performs binary PE classification, and future work could extend it to more granular diagnostic tasks, such as thrombus burden quantification, embolism subtype identification, or longitudinal progression analysis, which would further support quantitative and decision-oriented clinical applications. Notably, the proposed two-stage architecture is not inherently restricted to binary classification. By replacing the output layer and adopting multi-class or regression-based objectives, the dual-branch framework could be extended to support thrombus burden quantification or embolism subtype classification. Moreover, the frequency-aware enhancement strategy and dual-source feature fusion paradigm may also be applicable to other contrast-limited CT angiography tasks, such as low-dose vascular imaging in different anatomical regions. These potential extensions are beyond the scope of the current study and will be explored in future work. Finally, although the model demonstrates excellent quantitative performance, its interpretability requires further improvement. Techniques such as saliency mapping, attention visualization, or radiomic feature analysis could enhance clinicians’ understanding and trust in the model’s predictions.

In summary, this study demonstrates that a two-stage deep learning framework combining super-resolution enhancement with dual-branch classification enables robust and reliable PE detection under low-iodine, low-radiation CTPA conditions. The proposed approach not only improves image visualization quality but also leverages complementary feature information, offering a feasible pathway toward AI-assisted low-exposure, high-safety pulmonary embolism screening.

Future work will focus on several directions to further strengthen the clinical applicability of the proposed framework. First, multi-center validation will be conducted to evaluate model generalizability across different institutions, scanners, acquisition protocols, and patient populations, addressing potential domain shifts inherent in real-world clinical deployment. Second, we plan to explore integration of the framework into routine clinical workflows, such as PACS-based post-processing or real-time decision support during CTPA interpretation, with particular attention to inference efficiency and usability for radiologists. Third, improving model interpretability remains an important objective. Techniques such as attention map visualization, feature attribution analysis, and clinically meaningful region highlighting will be investigated to enhance transparency and increase clinician trust. In addition, the framework may be extended to more advanced diagnostic tasks, including thrombus burden quantification, embolism subtype classification, and longitudinal disease assessment, further supporting comprehensive and decision-oriented PE management.

Conclusion

This study proposes a novel two-stage deep learning framework specifically designed to address the challenges of pulmonary embolism (PE) detection under low-iodine, low-radiation CTPA imaging conditions. Unlike most existing PE detection methods that rely on standard-dose, high-contrast imaging, the proposed framework explicitly targets clinically constrained low-exposure scenarios, where reduced vascular enhancement and increased image noise substantially complicate diagnosis. By integrating a Frequency-Aware Super-Resolution Network (FASRN) to restore diagnostically critical vascular structures with a Dual-Branch Classification Network (DBCN) that jointly exploits complementary features from original and super-resolved images, the framework effectively balances image enhancement and diagnostic reliability.

Experimental results on real clinical low-exposure CTPA data demonstrate that the proposed method achieves strong and robust diagnostic performance, with an AUC of 0.928, sensitivity of 88.3%, specificity of 82.9%, and an F1 score of 79.5%, outperforming baseline models that use only original or SR-enhanced images. In addition, the framework maintains stable performance under noisy imaging conditions, highlighting its robustness and suitability for low-iodine, low-radiation clinical environments.

Beyond quantitative performance gains, this work has important implications for improving diagnostic safety in clinical practice. By enabling reliable PE detection while substantially reducing iodine contrast dosage and radiation exposure, the proposed framework supports safer CTPA imaging protocols, particularly for high-risk patient populations such as elderly individuals, patients with renal impairment, or critically ill patients. Its modular design further facilitates seamless integration into existing clinical imaging workflows, enhancing its practical applicability.

Overall, this study advances low-exposure PE detection by providing a clinically meaningful and technically robust AI-assisted solution that extends diagnostic capabilities beyond ideal imaging conditions. Future work will focus on multi-center validation, extending the framework to more advanced diagnostic tasks such as thrombus burden quantification and embolism subtype classification, and improving model interpretability to promote broader clinical adoption.

Author contributions

X.Z. organized the entire project and supervised the experimental design and methodological framework. M.H. is responsible for the main experiments, including image classification model development, training, evaluation, and manuscript writing. T.G. is responsible for clinical data acquisition, annotation, and coordination, ensuring high-quality low-iodine, low-radiation CTPA dataset. H.A. and X.F. are responsible for the design of the super-resolution model and related experiments. All authors reviewed and approved the final version of the manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (Grant No. 62521007, 62431011), the Fundamental Research Funds for the Central Universities (E2ET1104) and National High Level Clinical Research Funding (BJ-2025-213).

Data availibility

The present study used CTPA images from publicly available datasets RSNA as well as from Beijing Hospital. Public datasets can be accessed through the respective websites. In addition, the low-iodine, low-radiation CTPA dataset was collected from real clinical acquisitions at Beijing Hospital, with all procedures approved by the institutional ethics committee (Approval No. 2024BJYYEC-KY089-02). Patient data were fully anonymized to protect privacy. The clinical low-iodine, low-radiation CTPA dataset has been made available via a controlled-access repository and can be accessed at: https://pan.baidu.com/s/1PvfT_13mLlcHLLE-C5_g_g. Due to ethical and privacy considerations, access is granted for research and reproducibility purposes only. The access code can be obtained by contacting the corresponding author upon reasonable request.

Code availability

The implementation code for the proposed two-stage deep learning framework, including the Frequency-Aware Super-Resolution Network (FASRN) and Dual-Branch Classification Network (DBCN), will be released upon manuscript acceptance on a public repository. This will allow full reproducibility of the reported results and facilitate future research on low-exposure CTPA imaging.

Declarations

Competing interests

The authors declare no competing interests.

Consent to participate

The requirement for informed consent was waived by the Beijing Hospital Ethics Committee (2024BJYYEC-KY089-02) because this study was retrospective and used fully anonymized clinical data.