Automated Aortic Quantification Based on Artificial Intelligence: Validation Using Contrast-Enhanced and Non-Contrast CT Scans from the Same Session

1. Introduction

As the primary artery, the aorta is pivotal in transporting oxygen-rich blood from the heart to the rest of the body, ensuring body organs receive the necessary oxygen for survival[1]. Its structure is vital for the integrity of this major blood vessel, with issues such as aneurysms and dissections posing significant life-threatening risks[2,3]. Aortic dilation is a crucial precursor to the aortic aneurysm, dissection, or in patients with aortic valve disease, requiring treatment intervention once surpassing certain thresholds[4,5]. The prevalence of aortic dilation ranges from 1.4% to 23%, varying across different study populations[6,7,8]. Aortic dilation is related to ageing and existing cardiovascular diseases or abnormalities, leading to an elevated risk of subsequent aortic disease[9].

Advanced diagnostic techniques, such as echocardiography, MRI, and CT, enable early detection of aortic anomalies, ensuring timely treatment[10,11]. Following the guidelines, early surgical or pharmacological interventions have significantly enhanced patient prognosis, reducing the incidence of catastrophic outcomes[4,5]. Early detection and monitoring of aortic diseases can improve patient outcomes and minimize societal costs[12]. Non-contrast CT scans, such as those performed in lung cancer screening[13] as part of health screening programs, often include aortic imaging, offering an opportunity for early detection of aortic abnormalities. Maximizing the utility of any CT examination can significantly contribute to the early detection of diseases.

Artificial Intelligence (AI) techniques have rapidly advanced, enabling numerous automated analyses in medical imaging regarding cardiovascular imaging to become feasible[14]. In clinical AI applications, the variance in imaging techniques can pose challenges[15]. With its quantitative Hounsfield Units (HU), CT imaging is suitable for clinical AI applications. TotalSegmentator[16], introduced in early 2022, is a powerful AI model capable of segmenting various organs in CT images, advancing AI applications in CT imaging, including automated aortic segmentation[17,18,19,20,21,22,23,24,25,26]. While TotalSegmentator can automatically segment various tissues and organs in CT, a detailed and validated process for aortic quantification remains lacking. Should a method for quantifying the aorta be developed based on it and undergo proper validation, it would greatly benefit clinical diagnosis and monitoring.

Our research aims to develop and validate a fully automated aortic quantification process using TotalSegmentator as the core model. The workflow integrates automated processes and validates performance using paired CT images (contrast-enhanced and non-contrast) from the same patient within the same imaging session, ensuring clinical applicability.

2. Materials and Methods

2.1. Study Population

This study was a retrospective compilation of routine clinical imaging, receiving approval from the Institutional Review Board (IRB) of Cheng Hsin General Hospital (Approval No. [980] 111A-58). The IRB waived informed consent as all data were de-identified before analysis to ensure privacy, with no patient contact or intervention involved. All methods were performed in accordance with relevant guidelines and regulations, including those outlined by the Declaration of Helsinki and the IRB of Cheng Hsin General Hospital. A retrieval of 314 cases from the hospital’s database was conducted between October 2021 and May 2023. Figure 1 depicts the patient selection waterfall diagram. Within this cohort of 314 cases, 29 exhibited pronounced aortic dissections with discernible true and false lumens. These instances were excluded from our study due to the challenge of distinguishing such cases in CT without contrast. Furthermore, 95 cases were excluded as they only had contrast-enhanced CTs, which did not meet our inclusion criteria. Consequently, 190 cases (380 scans) were incorporated into the study, with an age range between 42-94 (median of 74) and a gender ratio of 98:92 (male to female). Of these, we conducted manual aortic annotations on 134 cases (Test Set 1), while the remaining 56 cases did not receive such annotations (Test Set 2). All 190 cases came with clinical records from physicians, who, within their clinical practice, only proceeded with measurements in the presence of an apparent arterial dilation.

2.2. CT Imaging Protocol

The CT images were acquired using a dual-source 128-slice multidetector CT scanner (Somatom Definition FLASH, Siemens Healthineers, Erlangen, Germany). Two distinct imaging protocols were implemented. For contrast-enhanced scans, a slice thickness of 1 mm was selected, whilst non-contrast scans were conducted with a 3 mm slice thickness. All scans were performed at a voltage of 120 kVp with automated tube current modulation (CARE Dose4D, Siemens Healthineers). Additionally, pixel spacing for both protocols was the same at 0.65 mm, ensuring consistent spatial resolution across images. During imaging, patients were instructed to fully inhale, hold their breath, and remain still until advised to breathe normally again.

2.3. Ground-Truth Segmentation

We used the open-source medical image analysis platform 3D Slicer (https://slicer.org/) for manual semi-automated aorta annotation on contrast-enhanced CT images, utilizing the SegmentEditorExtraEffects and SlicerVMTK extensions. The watershed algorithm in the former segmented the aorta based on local histograms, while the Extract Centerline tool automatically detected inlets and outlets to extract the vessel’s centerline. After initial segmentation, inaccurately segmented sections were corrected in 3D space, and the central line of the aorta was calculated. The aorta was then divided into the ascending aorta, aortic arch, and descending aorta based on the brachiocephalic trunk and left subclavian artery positions[27]. Annotations were performed by medical students and verified by experienced medical professionals.

2.4. Automated Aorta Quantification Workflow

Our automated aortic quantification process involved several steps (Figure 2). First, TotalSegmentator segmented tissues and organs in the CT images. The aorta was then extracted, and inlet and outlet points were automatically detected. If multiple points were identified, the two farthest apart were selected. Next, the vascular centerline was computed, straightened, and resampled to a uniform voxel size of 1 mm³. Finally, the aorta's maximal diameter and cross-sectional area were calculated for each segment using the resampled images.

We define segment levels based on the labelled positions, employing these distinct segments for subsequent AI results assessment. Additionally, leveraging the NumPy package (version 1.22.3), we automate the calculation of the maximum diameter and area size of cross-sections (Figure 3). A smoothing process was applied for the sequence of diameter and area values obtained using the Savitzky-Golay filter from the SciPy package (version 1.10.0), with a window length of 11 and a polynomial order of 2.

We developed an in-house 3D Slicer module, DeepAorta, integrating features from multiple extensions: TotalSegmentator (dcfa71b), SlicerVMTK (144582e), and Sandbox (94e2df0). The Extract Centerline tool in SlicerVMTK enabled automatic detection of the aorta’s inlet and outlet points and centerline calculation, while the Sandbox Curved Planar Reformat extension facilitated vessel straightening and resampling based on the centerline. The software was deployed on 3D Slicer version 5.7.0, operating on a Windows 11 system. The primary hardware utilized was a CPU i9-13900K, with 64GB RAM, and a GPU NVIDIA GeForce RTX 4090 with 24GB VRAM.

2.5. Assessment of Measurements and Statistical Analysis

As previously mentioned, contrast-enhanced CT served as the gold standard for segmentation. We then compared the outcomes of AI-driven automatic quantification on contrast-enhanced and non-contrast CTs. Overlap-based metrics, such as the Dice similarity coefficient and the Intersection over Union (IoU), were evaluated alongside boundary-distance-based metrics, including the 95% Hausdorff distance and the average symmetric surface distance[28].

Furthermore, we evaluated the correlation and consistency for size measurements across the datasets. Our analysis employed the Pearson and intraclass correlation coefficients (ICC). This involves statistical analysis of aortic dimensions, including the diameter and area. Lastly, our study utilized Receiver Operating Characteristic (ROC) analysis between records in routine clinical procedures and AI quantification in identifying potential aortic dilation. Specific dilation thresholds were set for different regions[29]: 40mm for the ascending aorta, 35mm for the aortic arch, and 30mm for the descending aorta.

3. Results

3.1. Aortic Quantification Measurements

Table 1 presents the quantified results of aortic dimensions using GT, AI inference on CT with contrast and without contrast, divided into different segments: the entire aorta, ascending aorta, aortic arch, and descending aorta. The data encompassed maximum, mean, and median diameters/areas for the whole aorta, ascending aorta, aortic arch, and descending aorta.

Under the GT analysis, the whole aorta’s maximum diameter was 38.28±5.42 mm, with the mean and median diameters at 27.17±3.17 mm and 25.31±3.14 mm, respectively. The maximum area recorded was 1047.94±350.20 mm², with mean and median areas at 554.99±136.80 mm² and 461.40±112.76 mm². An increase in aortic dimensions was observed utilizing AI inference on CT with contrast. The whole aorta’s maximum diameter increased to 40.62±5.29 mm, and the maximum area to 1212.65±357.18 mm². Similar increments were noted across the ascending, arch, and descending segments. AI inference on CT without contrast further escalated the dimensions, where the whole aorta’s maximum diameter reached 42.29±5.32 mm, and the maximum area was 1289.10±342.76 mm².

3.2. Segmentation Metrics

Table 2 delineates the segmentation metrics between GT and AI inference in CT images (CT_AI), both with and without contrast.

For GT versus CT_AI with contrast, the Dice coefficient ranged from 0.86±0.02 in the descending aorta to 0.91±0.02 in both the ascending aorta and aortic arch. The IoU scores were lowest for the descending aorta at 0.76±0.03 and highest for the ascending aorta at 0.84±0.03. The 95% Hausdorff distance was lowest for the ascending aorta at 2.27±0.53 mm and highest for the descending aorta at 2.63±0.56 mm. The average symmetric surface distance varied minimally, with the lowest observed in the aortic arch at 0.91±0.17 mm and the highest in the descending aorta at 1.44±0.17 mm.

In the comparison between GT and CT_AI without contrast, the Dice coefficients indicated lower conformity than CT_AI with contrast, ranging from 0.78±0.04 in the descending aorta to 0.87±0.03 in the ascending aorta. The IoU scores mirrored this trend, with the lowest score for the descending aorta at 0.65±0.05 and the highest for the ascending aorta at 0.77±0.05. The 95% Hausdorff distance increased, with the lowest distance in the aortic arch at 3.64±1.05 mm and the highest in the whole aorta at 4.24±0.72 mm. The average symmetric surface distance similarly increased, with the lowest in the aortic arch at 1.45±0.51 mm and the highest in the descending aorta at 2.37±0.35 mm.

3.3. Correlation and Consistency Analysis

Figure 4 and Figure 5 depict the correlation coefficients and ICCs for the maximum diameter, illustrating the comparison between GT and CT_AI. (Please refer to Tables S1 and S2 for comprehensive values.)

In CT_AI with contrast, there was a high degree of similarity to the GT across all metrics. The correlation coefficient for the maximum diameter of ascending aorta was 0.987 (95% CI: 0.98-0.99). A similar pattern in ICCs, consistently near 1, indicated high reliability in measurements between GT and CT_AI with contrast.

The analysis of CT_AI without contrast revealed a decline in correlation coefficients and ICCs. The correlation coefficient for the maximum diameter of ascending aorta decreased to 0.945 (95% CI: 0.92-0.96), with a similarly reduced ICC. This trend was accentuated in the aortic arch, where the lowest ICCs were observed. This segment demonstrated the most variation, with correlation coefficients for median diameter dropping as low as 0.678 and ICCs reflecting this reduced agreement (Table 3 and Table 4). The descending aorta, while showing higher overall agreement than the arch in the CT_AI without contrast, presented a reduction in correlation coefficients and ICCs compared to the CT_AI with contrast.

3.4. Receiver Operating Characteristic Analysis

In the ROC analysis (Figure 6), we compared the efficacy of physicians’ routine clinical records and automated quantification in detecting aortic dilatation. Test Set 1 included 134 cases with GT segmentation, and Test Set 2 contained 56 cases lacking GT. These 190 cases featured routine clinical records from physicians, who recorded dilation measurements in the ascending aorta noted during clinical routines. For Test Set 2, we used the CT_AI’s quantification results from contrast-enhanced CTs, adjusted via regression correction from Test Set 1, as GT measurements. We were comparing the efficacy of physicians and CT_AI in detecting aorta dilation in images without the contrast agent.

For the ascending aorta in Test Set 1, the physician achieved an AUC of 0.88 [0.80-0.95], which was surpassed by the CT_AI. CT_AI with contrast reached the maximum AUC value of 0.997 [0.99-1.00], while that on CT_AI without contrast also demonstrated superior accuracy with an AUC of 0.984 [0.96-1.00].

Considering the aortic arch in Test Set 1, there were no clinical records and CT_AI with contrast had an AUC of 0.976 [0.95-0.99], whereas the performance of CT_AI without contrast declined with an AUC of 0.874 [0.80-0.93]. For the descending aorta in Test Set 1, there were no clinical records, and the CT_AI maintained a high level of diagnostic accuracy with contrast (AUC = 0.990 [0.97-1.00]) and without contrast (AUC = 0.942 [0.90-0.97]).

In Test Set 2, which lacked the GT segmentation and involved CT_AI with contrast by GT correction as GT measurement for the aorta, CT_AI performed robustly without contrast, achieving an AUC of 0.966 [0.89-1.00]. The physician’s AUC in this set was 0.841 [0.73-0.94], lower than the CT_AI without contrast. Table 5 documents the optimum sensitivity and specificity within the ROC analysis.

4. Discussion

We developed a fully automated aortic quantification process using the open-source model and tools, demonstrating high effectiveness in detecting aortic dilation across CT imaging protocols, both with and without a contrast agent. The high degree of correlation and consistency in aortic quantification metrics (close to 1 for the maximum), alongside the impressive AUC scores in detecting vascular dilation (exceeding 0.9 across various segments and imaging protocols), underscore this fact. This reflects the model’s potential to facilitate the early detection and longitudinal monitoring of aortic pathologies. This has the potential to redefine routine clinical imaging processes and, by promoting the early detection of aortic anomalies, to improve the prognoses for affected patients.

Our study observed distinct performance in evaluating aortic segments—ascending aorta, aortic arch, and descending aorta. Our semi-automatic method was based on contrast thresholds that defined the aortic lumen as GT. The AI’s higher measurements stemmed from including the aortic wall, increasing thickness akin to the wall size[30]. The lack of wall contrast in non-contrast CT led to an even larger inclusion area. Nonetheless, the correlation and consistency results demonstrated the AI’s exceptional capability in delineating the ascending aorta through contrast-enhanced and non-contrast CTs. Besides, AI performed better in contrast-enhanced CT than non-contrast CT, likely because the ground truth was based on contrast-enhanced CT, and the contrast agent highlighted vascular boundaries, improving AI segmentation and quantification. Irrespective of using a contrast agent, the AUC scores for all three aortic segments consistently exceeded 0.9 and surpassed clinical records, accentuating the AI’s heightened sensitivity in detecting aortic dilation abnormalities. Nevertheless, despite a slight decline, the AI performance on non-contrast CT remains commendable. This suggests that AI could be utilized in scenarios where contrast agents are contraindicated, thereby widening the scope of non-invasive cardiovascular diagnostics. The capacity to discern aortic structure without contrast may have significant implications for patients with renal impairment or allergies to contrast media, offering a safer diagnostic alternative.

Clinical quantification of the aorta requires specialized software, and manual measurement is time-consuming and subject to certain errors[26]. If the imaging is not specifically conducted to examine the aorta, measurements are only undertaken when a noticeable dilation is observed upon reviewing the images. The AI application for automated aorta quantification saves time and keenly assists and alerts physicians to potential aortic abnormalities in patients. Recent research focusing on AI’s application and validation for the automated quantification of the aorta has verified the exceptional accuracy of AI in aortic measurements, demonstrating its feasibility for clinical application[17,18,19,20,26]. Additionally, AI has proven adept at identifying aortic pathologies like aneurysms and dissections[22,23,24,25], with sensitivity and specificity on par with physicians, further highlighting its potential in clinical usage. By evaluating the AI performance across contrast-enhanced and non-contrast CTs in identical cases, our research has leveraged paired imaging to affirm AI’s versatility in non-contrast CT settings, contributing to the discipline. Moreover, in contrast to other non-open source or commercial models, our approach is based entirely on the open-source model and tools, enhancing the reproducibility of our proposed methodologies.

Our study is subject to some limitations. Firstly, our validation was conducted on a single machine within a single center. However, the open-source core model we utilized, though not explicitly designed for aortic quantification, has been validated across multiple centers and machines for its accuracy and stability¹⁶. Secondly, the clinical records of aortic dilation measurements performed by various physicians were based on subjective assessments, with measurements taken only when aortic dilation was suspected. However, we addressed this issue by delineating the entire aorta and obtaining GT measurements through image processing. Theoretically, this approach should be more accurate and sensitive, a fact that was corroborated by our ROC analysis. Thirdly, our proposed process, centered around the TotalSegmentator as the core model, does not automatically segment the aorta into specific sections or infer results for different aortic landmarks as recommended[4,5]. Instead, our approach relies on statistical methods to identify the position of maximum values and calculate averages and medians from the entire aorta and different segments. Despite these constraints, we have validated the core model’s accuracy in aorta quantification. Future developments based on this model and methodology may train more robust AI systems capable of automatically segmenting distinct aortic sections and detecting landmarks of interest.

5. Conclusions

We have developed an automated aortic quantification method anchored in the open-source model and tools. Our validation reveals that AI-powered automated aortic quantification enhances the clinical detection of abnormal aortic dilation via CT imaging, offering greater sensitivity than clinical routine records. This technology, which can be used in contrast-enhanced and non-contrast CT, affords physicians critical support in the early identification of unusual disease progression and the vigilant monitoring of patient aortic health.