3D printed flexible anatomical models for left atrial appendage closure planning and comparison of deep learning against radiologist image segmentation

doi:10.21203/rs.3.rs-2188108/v1

Download PDF

Research Article

3D printed flexible anatomical models for left atrial appendage closure planning and comparison of deep learning against radiologist image segmentation

https://doi.org/10.21203/rs.3.rs-2188108/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background:

Medical 3D printing is being increasingly employed for pre-procedural planning and simulation. One important application is in occluder device sizing for left atrial appendage (LAA) closure. Studies have demonstrated clinical utility of 3D printed anatomical models for LAA closure. Artificial intelligence-based segmentation has been applied to multiple cardiovascular diseases, including to LAA segmentation. However, to our knowledge, there has not been a comparison of artificial intelligence-based deep learning segmentation (DLS) where there was a clinical 3D printed model of the left atrium and appendage.

Methods:

Thirty-nine patients had 3D printed models requested by the interventional cardiologist (IC). Standard segmentation (SS) was performed by a trained engineer and approved by a cardiovascular imager (CI). The models were 3D printed using flexible resin and desktop inverted vat photopolymerization technology. The effort expended throughout the workflow was meticulously documented. Thirty-four of the 39 patients underwent left atrial appendage occlusion using the 3D printed model for device sizing. The 34 patients who underwent a procedure using the 3D printed model were followed for 6 months for major adverse events, device embolization, procedure related myocardial infarction (MI), procedural stroke, new pericardial effusion, pericardial effusion requiring intervention, surgical conversion, and peri-procedural death. All 39 patients also underwent DLS using a commercial software and metrics including segmentation time, segmented volume, DICE index were assessed compared to the SS. A Bland-Altman and regression/correlation analysis was also conducted.

Results:

The mean SS time was 72.3 minutes whereas the mean DLS time was 3.5 minutes. The DLS closely matched the SS with a mean DICE index of 0.96. The average number of devices attempted was 1.3. The DLS was highly correlated with the SS volume data (r = 0.99). Bland-Altman analysis showed a negative bias (-5.31%) in the volume difference data. There were no long-term complications in the 34 patients who underwent procedure using the 3D printed model for occluder device sizing.

Conclusions:

We have successfully demonstrated the performance of a commercial DLS algorithm compared to CI approved SS for left atrial appendage occluder device sizing using a clinical 3D printed model.

Medical 3D printing

deep learning

artificial intelligence

image segmentation

left atrial appendage closure

medical device

vat photopolymerization

additive manufacturing

rapid prototyping

In the United States, 3–6 million people have atrial fibrillation (AF) and the disease prevalence has been estimated to climb to 6–16 million by 2050 [1]. AF predisposes to thromboembolic events, and mechanical occlusion of the left atrial appendage (LAA) in patients with contraindications or intolerance to oral anticoagulation, is an alternative strategy for stroke prevention [2]. Trans-esophageal echocardiography (TEE) is the current gold standard for pre-procedural imaging, although cardiac computer tomography (CT) offers improved imaging with high-quality multiplanar 3D reconstructed images [3], and is associated with a higher device implantation rate, shorter procedural time, and less frequent change of devices [4]. Because LAA morphology is variable, it can be intuitively challenging for the interventional cardiologist (IC), even with preprocedural two-dimensional (2D) and volumetric imaging to accurately size a closure device [5]. Improper evaluation of the morphology can result in procedural failure [6].

Three-dimensional (3D) printing provides tactile feedback for complex surgical planning [7]. For complex interventions, 3D printing is used complementary to 3D visualization, defined as the collective volumetric representations of volumetric anatomy displayed on a 2D screen [8]. Since image segmentation is similar for both 3D visualization and 3D printing, comparisons between these two methods of intervention planning are reasonable and can help identify clinical scenarios for which the additional work of 3D printing adds value with respect to 3D visualization alone. Moreover, considering all of the steps for creating an anatomic model, segmentation is prone to human inaccuracy [9], and for this reason quality control includes verifying image segmentation [10, 11], commonly employing the residual volume technique [12].

Anatomic models are increasingly used for cardiac procedures [13], including 3D printing of the left atrium (LA) and LAA for LAA occlusion device sizing [6]. The in vitro simulation of occlusion using 3D printed models was found to result in a single successful occlusion procedure without surgical complications [14]. On a larger scale, a meta-analysis and systematic review found that the pre-procedural device sizing showed better agreement with the implanted device compared to the sizing based on TEE or CT [15]. Anatomic models 3D printed in a compliant (“rubber-like”) material have been useful in patients where the anatomy is complex and the interaction between the device and LAA is difficult to quantify [6]. Concordance between the 3D printed models and inserted Amplatzer devices has been reported, although the agreement between the 3D printed models and Watchman devices was poor in the same study [16]. Two studies reported that the 3D printed model correctly predicted the size of inserted device in all 8 and 10 patients, respectively [17, 18]. Another study reported that 3D printed anatomic models can be used to identify the device size, implantation depth, and orientation [19]. Three-dimensional printed models that included the LA, LAA, aortic annulus, and most proximal portions of the superior and inferior vena cava were found to be useful in helping simulate device deployment using different catheter configurations [20]. A study reported more consistent measurements of the ostium area and LAA volume using virtual 3D models compared to the reference standard 2D measurements [21].

Artificial Intelligence (AI) has made important contributions to vision and image processing research, and has now been applied for medical imaging segmentation [22–25]. If proven accurate for specific clinical scenarios – or as a more general segmentation tool, it would be well suited for 3D printing because of the labor-intensive nature of segmentation. A joint-atlas-optimization was developed to segment the LA, pulmonic veins (PVs) and LAA from magnetic resonance angiography images and showed a significantly higher dice coefficient in the segmented PVs (small distant part in LA geometry) and LAA, although the resulting segmented model was not employed clinically [26]. A 16-layer convolutional neural network was developed for segmenting the LA epicardium and endocardium resulting in the highest dice coefficient indicating strong potential for application to AF treatment [27]. A 3D U-net with contour loss was employed to segment the LA resulting in high dice coefficient, specificity and sensitivity [28]. A U-net integrated with Kalman Filter was developed for LA segmentation and yielded high dice coefficient when retrospectively compared to radiologist ground truth segmentation in 20 patients [29]. However, none of these studies had a verified reference standard that includes a 3D printed anatomic model that was approved by a cardiovascular imager (CI), used for procedural planning, validated as accurate by the proceduralist using data collected during the deployment of the occluding device, and then further validated by long-term clinical outcomes post-procedure.

Segmentation studies are challenging because it is difficult to establish a reference standard. One general strategy is to test segmentation of known volumes such as a phantom or cadaveric material. However, this does not account for human physiology, and this approach is very limited for cardiac applications [30, 31] due to errors in segmentation amongst other not accounting for heart motion[32] For cardiac applications, the most valuable reference standard segmentations are arguably those performed by or verified by a physician who is board certified and licensed to perform this clinical service.

This project uses a complete sub-study to establish an optimized reference standard from which a commercial AI-based segmentation can be tested. The sub-study includes a cohort of consecutive patients for which the segmentation was not only performed and/or checked by an attending

CI, but also verified after the procedure by the IC. Moreover, the patients that formed the reference standard cohort underwent extensive clinical follow-up to ensure that there were no adverse outcomes that could in theory be ascribed to the 3D printed model, including the segmentation. Once the reference standard was secure via the sub-study, the purpose of this study was to describe the characteristics of an AI-based segmentation tool for medical 3D printing, and to test the hypothesis that AI-based segmentation is accurate using as reference standard the segmentation of a cohort of clinical LAA anatomic models used for planning LAA occlusion with a Legacy Watchman device using a flexible 3D printed anatomic model.

The study was conducted in two parts (Fig. 1). Part I (boxed red) involved the 3D printing for all patients and the clinical follow-up of a subset of those patients who underwent procedure. Part II of the study included Deep Learning Segmentation (DLS), comparison against Standard Segmentation (SS), and data/statistical analysis.

Part 1: Clinical sub-study

Patients

The University of Cincinnati Institutional Review Board approved this retrospective, observational study. The patient cohort was a consecutive set of patients between February 2020 and May 2021 with contraindication to oral anticoagulation for whom a 3D printed anatomical model was requested for LAA closure at the University of Cincinnati Medical Center, Cincinnati, OH, USA.

Clinical Cardiac CT

Axial CT (256 x 0.625mm detector row, 280 millisecond gantry rotation time) data was acquired with prospective Electrocardiogram (ECG) gating after the administration of 80–100 mL iodinated contrast material. The current and voltage used for the CT acquisition were prescribed by the attending CI, based on typical factors for CT such as renal function and patient body habitus. Images were reconstructed at 0.625–1.25 mm slices without overlap and archived in the hospital Picture Archiving and Communication System (PACS).

Images for all patients were reviewed by a CI with 20 years of experience in a dedicated cardiac CT practice and included multiplanar reformation (MPR) and volume rendering of the LAA morphology [33, 34] with specific measurements of the LAA orifice orthogonal to its plane (Fig. 2). The appendage morphology was recorded from the cardiovascular imaging report following the 4 types commonly reported in the literature – chicken wing, cactus, windsock, and cauliflower.

3D Printing Standard Segmentation (SS)

The LA, LAA, and proximal segments of the PV underwent initial segmentation from the thin section images by a 3D printing engineer on a dedicated workstation (InPrint 3.0, Materialise, Leuven, Belgium) using a combination of auto-segmentation and direct planimetry [35–37]. The segmentation included an initial thresholding of the blood pool, after which the LA was cropped on the axial and coronal projections. A manual split mask excluded those voxels outside the LA and LAA that were inadvertently included during thresholding. The proximal segments of the PVs were included, with special attention directed for those patients who did not have 2 left and 2 right pulmonary veins. This segmentation was then checked for accuracy and further manipulated using direct planimetry by a CI. The final segmentation was approved by the attending CI, and this was determined to be the reference standard and referred to as “Standard Segmentation” (SS).

Computer Aided Design and 3D Printing

A 2mm wall was defined and built around the segmented volume. While preserving all PVs, a large portion of the LA was digitally removed (Fig. 3) near the mitral valve region to facilitate removal of support material (Fig. 4), particularly inside the appendage. This also enabled monolithic 3D printing using desktop inverted vat polymerization (VP). An additional smoothing operation was performed in Materialise 3-matic (Materialise, Leuven, Belgium) followed by labeling of the model with the patient’s medical record number [8]. The final STL file was 3D printed with a Formlabs Form 3 VP 3D printer (Formlabs, Sommerville, MA, USA) using Elastic resin (Formlabs, Sommerville, MA, USA) [38] to better emulate cardiac tissues when compared to harder resins (Fig. 5). The PVs were trimmed and cut open using Materialise InPrint 3.0 (Materialise, Leuven, Belgium). The support tip size was set to 0.4mm and the support density to 0.90 in PreForm software (Formlabs, Sommerville, MA, USA). The layer height used was 0.1mm. Post printing, the models were carefully detached from the build plate and rinsed in 91% isopropyl alcohol in a Formwash unit (Formlabs, Sommerville, MA, USA). The support structures were gently removed, and the model was UV cured in a Form Cure unit (Formlabs, Sommerville, MA, USA) for 20–25 min at 65 degrees Celsius.

Watchman Device Procedural Planning and Deployment

Both the CT report and the 3D printed models were used for surgical planning. In all patients, the Legacy Watchman™ device (sizes 21mm, 24mm, 27mm, 30mm, or 33mm; Boston Scientific, Marlborough, MA) was deployed. As part of the clinical sub-study that was used to establish the reference standard, intervention planning with the anatomic model was considered superior to 3D visualization for all patients as reported by the single IC who performed all the procedures. Moreover, for all patients, the 3D printed model was considered superior to using 3D visualization alone. The rationale for that decision was that the 3D printed model was used in a pre-procedural conference that included the IC, a representative from Boston Scientific, the echocardiographer, and the anesthesiologist. During that conference, a nonsterile Watchman device was deployed into the model (Fig. 6) to assess size and angulation between the device and the LAA orifice. 3D visualization measurements were also consulted and correlated to those made directly on the model. The final plan was determined by consensus, based on the available device sizes, and the shape and dimensions of the LAA.

The patient was placed under general anesthesia, and TEE was performed by the echocardiographer to guide device deployment. A femoral vein was accessed through which transseptal puncture was performed under TEE guidance. Once LA access was obtained, the patient was maintained at an active clotting time of greater than 300 seconds via sequential heparin boluses. Before deployment, angiography of the LAA was performed using a RAO caudal view in keeping with the planning derived from the anatomic models. After fluoroscopy was adjusted to eliminate projectional shortening, the dimensions of the LAA were confirmed against the anatomic model. The deployment fluoroscopic angles were also confirmed, as were the contours of the segmented volume. This ensured that the device size correlated with the pre-procedural planning. The Watchman device was then released, after which stability, placement, and perfect seal were assessed with both fluoroscopy and TEE.

Patient Clinical Follow Up

All patients underwent clinical follow-up for a minimum of 6 months. The complications assessed included major adverse events, device embolization, procedure related myocardial infarction (MI), procedural stroke, new pericardial effusion, pericardial effusion requiring intervention, surgical conversion, and peri-procedural death. All complications were adjudicated by the staff performing the procedure and the imaging specialists, with attention paid to whether or not the complication could have been related globally to the use of the anatomic model for procedural planning, and if so, could the complication been related to the image segmentation.

Part 2. Study of AI based segmentation for LAA occlusion planning

Deep Learning Segmentation (DLS)

Independent from and after the completion of data collection for the clinical sub-study, the same Digital Imaging and Communications in Medicine (DICOM) data set for each patient was segmented using the AI segmentation online tool CT Heart in the Mimics Viewer (Materialise, Leuven, Belgium). The tool is proprietary, and it is not cleared by the United States Food and Drug Administration for an intended use.

The accuracy of DLS was not assessed by the CI; the algorithm was executed entirely for research purposes. The details of the segmentation (4 steps below) were provided by the vendor. The algorithm expects a cardiac scan with contrast as input and computes the masks for the following structures: aorta (AO), pulmonary artery (PA), left ventricle (LV), right ventricle (RV), LA, right atrium (RA), and myocardium (MYO). For this project, the only region of interest was the LAA, the LAA, and the proximal PVs. The final segmentation output was referred to as the deep learning segmentation (DLS). It includes these steps:

Heart Localization: To deal with different fields of view and to focus on the heart, a localization step that retrieves a bounding box that includes the heart and part of the descending aorta and spine is used. This step is based on a deep learning two-dimensional (2D) localization algorithm called Faster R-CNN [39], which is applied to the projections of the input image in the different axis to identify a bounding box around the heart.

Segmentation initialization step: With the region of interest detected in the localization block, the popular segmentation algorithm for medical imaging U-Net is employed to initialize the segmentation of the structures of interest [40]. The initialization step uses a small training set to train the U-Net model resulting in a rough segmentation of the heart which is finetuned in the subsequent algorithmic steps.

Seed and bloodpool generation: The mask obtained from the previous step is then used to extract a representative thresholding value for the right and left heart to create a bloodpool for each side as well as to generate a limited set of seed points for each structure (AO, PA, LV, RV, LA with LAA and PVs, RA, and MYO).

Split mask and final output: Using the split mask tool included in Mimics Innovation Suite (Materialise, Leuven, Belgium), the bloodpool (left and right heart) is split into the desired individual anatomical structures. After the deep learning algorithm segmented the volume, the distal portion of the pulmonic vein including its smaller branches were trimmed manually to match the SS and allow fair comparison of the two volumes.

Test Metrics for the AI Segmentation

For all patients, the clinical service 3D printed an anatomic model, and the model was printed from the SS approved by the CI. The DLS was registered against the SS using Materialise 3-matic 15.0 (Materialise, Leuven, Belgium). The following metrics [41] were used to compare the DLS versus the SS: Dice similarity coefficient (DSC), the average symmetric surface distance (ASSD) and the Hausdorff distance (HD) for each registered (DLS, SS) pair.

Statistical Analysis

Statistical analyses were performed with GraphPad Prism Software V9.4.1 (San Diego, CA, USA). Results are expressed as mean ± standard deviation (SD) or as median and interquartile range (IQR) for continuous variables, and as absolute counts and percentages for categorical variables. Paired t-tests were used to evaluate statistical significance between the DLS and SS volume and segmentation time data. A p-value of ≤ 0.05 was considered as the level of statistical significance. Agreement analyses comparing the SS and the DLS was performed using Bland-Altman analysis to assess the bias and the limits of agreement. A regression analysis was performed and correlation was also assessed.

Patients

A 3D printed anatomic model was requested for 39 patients (Table 1) with AF who were considered for LAA occlusion device placement between December 2019 and June 2021. Among these 39 patients, 34 patients underwent the procedure.

Table 1

Baseline patient characteristics of the 34 patients who underwent Legacy Watchman™ device Occlusion.
Age, Years	73.5 ± 6.9
Sex, Male	19 (55.9)
BMI, kg/m²	31.7 ± 7.0
CHADS₂-VASc Score	4.9 ± 1.6
Has-BLED Score	3.4 ± 0.7
History of CHF	10 (29.4)
History of Hypertension	33 (97.1)
History of DM	16 (47.1)
History of Stroke/TIA	13 (38.2)
History of prior bleeding	24 (70.6)
History of ICH	6 (17.6)
History of gastrointestinal bleed	12 (35.3)
History of renal impairment	10 (29.4)
LVEF, %	56 ± 12
Concomitant procedures	8 (23.5)
Baseline Creatinine, mg/dL	1.18 ± 1.29
Maximum ostium diameter (intraprocedural TEE), mm	19.9 ± 3.3*
Data presented as mean ± SD or n (%). Abbreviations are as follows: BMI, body mass index; CHF, congestive heart failure; DM, diabetes mellitus; ICH, intracranial hemorrhage; LVEF, left ventricular ejection fraction; TEE, transesophageal echocardiogram; and TIA, transient ischemic attack.
* Maximum ostium diameter data was not available in 5 patients.

Table 2

Procedural outcomes and complications in the 34 patients who underwent Legacy Watchman™ device Occlusion.
Successful device implantation	34 (100%)
Major adverse events	GI Bleed 3 (8.8%)
Device embolization	0 (0)
Procedure related MI	0 (0)
Procedural stroke	0 (0)
New pericardial effusion	Trivial 2 (5.8%); Small 1 (2.9%)
Pericardial effusion requiring intervention	Moderate and requiring drainage 1 (2.9%)
Surgical conversion	0 (0)
Peri-procedural death	0 (0)
Data are presented as n (%). Abbreviations are as follows: MI, myocardial infarction; GI, gastrointestinal.

Clinical Cardiac CT

All 39 patients successfully underwent cardiac CT and there were no image artifacts that precluded image interpretation or image segmentation. There were no artifacts that resulted in a limitation in the generation of 3D printed anatomic models. The LAA morphology distribution was as follows: 5 Cactus (12.8%), 24 Chicken Wing (61.5%), 5 Cauliflower (12.8%), and 5 Windsock (12.8%).

3D Printing Standard Segmentation (SS)

The image segmentation for all 39 patient DICOM files who underwent 3D printing was approved by a CI. All image segmentations underwent accuracy check by a CI, and none of the models had a noticeable error (using CT images as reference standard). The mean ± SD segmentation time as recorded by the engineer who performed the preliminary segmentation was 72.3 ± 40.4 min (range 30–150 min).

Computer Aided Design and 3D Printing of Models

All 39 models were 3D printed successfully using Elastic resin (Fig. 5) with complete removal of all internal support structures. The average CAD time across all patients was 38.5 ± 13.7 min (range 30–60 min). The mean ± standard deviation print time was 18.1 ± 5.8 hours (8–30 hours) and material usage was 78.3 ± 18.9 ml (range 51.4–124.9 ml). The 2mm anatomic model walls were considered appropriate for both sizing and test deployment of the Watchman device (Fig. 6).

Watchman Device Procedural Planning and Deployment

The number of devices attempted was 1.3 ± 0.9 (1 device in 28 patients, 2 devices in 4 patients, 4 devices in 1 patient, 5 devices in 1 patient) across the 34 patients. The fluoroscopy time was documented in 19 and had a mean ± SD of 8.0 ± 5.8 minutes (range 2.3–25.8 minutes). The procedure time was documented for all patients had a mean ± SD of 85.5 ± 55.4 minutes (range 31.0–262.0 minutes).

Patient Clinical Follow Up

All 34 patients had successful LAA closure with the Legacy Watchman™ device. There were 3 periprocedural and 0 long term complications (Table 2). The 3 patients who had a GI bleed were all treated successfully, as were those with a pericardial effusion. None of the complications, with specific attention to the one patient who required a pericardial drainage, were attributed to the global use of the anatomic model.

Deep Learning Segmentation (DLS)

As noted, the DLS was performed outside of the clinical sub-study. The algorithm successfully converged for all patients, resulting in segmentation of the LA, the LAA, and the PVs (Fig. 7 and Fig. 8) in all 39 patients. The mean ± SD DLS time was 210.6 ± 12 sec (range 186–238 sec). The DLS time was significantly lower than the SS time (p < 0.0001).

Test Metrics

For all 39 patients who underwent 3D printing, the registration between the SS and DLS differed by no more than 0.5mm, on average. There was significant difference in the volume of the DLS blood pool and the SS (p < 0.05). There was a broad range (80.5 to 285.2 cc from SS) of LA blood pool volume. The mean DSC was 0.96 (range 0.92–0.99), reflecting that the DLS matched the SS across the entire range of volumes. Using the SS as reference standard, the greatest segmentation errors in the DLS algorithm occurred near the mitral annulus and where the PVs were truncated (Fig. 8).

The relative volume error was 5.74% with a similar median value of 5.23% across the 39 patients, demonstrating reasonably good agreement between the SS and DLS volumes (Table 3). The relative error was larger than 10% in 5 out of the 39 patients. Using the Bland-Altman method, the percent difference bias was − 5.31% (SD 4.10%) and the 95% limits of agreements bounds were from − 13.34–2.73% in the data comparing the DLS to the SS (Fig. 9). The DSC were high with an average of 0.96 and a range between 0.921–0.989 across the 39 patients. The average ASSD was 0.61mm and the average HD was 8.80mm (Table 4). The DLS and SS volume data were highly correlated (r = 0.99) and a linear regression analysis demonstrated good fit (R² = 0.98) (Fig. 10).

Table 3

Summary of the blood pool volumes and errors for the SS and DLS.
	Standard Segmentation (SS) Volume (cc)	Deep Learning Segmentation (DLS) Volume (cc)	Relative Error (%)
Maximum	285.2	278.1	12.05
Minimum	80.5	77.7	1.47
Average	170.6	162.3	5.74
Standard Deviation	49.5	49.1	2.86
Median	172.1	162.3	5.23

Table 4

Summary of the metrics comparing the SS and DLS.
	DSC	ASSD (mm)	HD (mm)
Maximum	0.989	1.84	19.32
Minimum	0.921	0.21	3.96
Average	0.960	0.61	8.80
Standard Deviation	0.017	0.34	4.51
Median	0.963	0.57	6.90

The present study reports the validation of a novel AI-based segmentation algorithm using as basis the radiologist approved segmentation that resulted in a 3D printed model used to plan the occlusion. Based on real world clinical data and utilizing as ground truth manually segmented anatomical structures, we demonstrated that a deep learning-based approach provides accurate models while substantially reducing processing time. Importantly, the reference models were employed in 34 patients illustrating their clinical value and the potential advantage, evident by the lack of events peri-procedurally as well as in the long-term follow-up.

A highly accurate automated image segmentation algorithm has the potential to reduce the time of image segmentation. If such algorithms were available and accurate, they could reduce the cost for medical 3D printing [42], given that the cost remains a major barrier to the wider adoption of this technology [43]. Regarding image segmentation, although there is a growing literature for AI-based cardiac segmentation, to our knowledge, there is no reported AI-based segmentation for LAA planning that used a 3D printed anatomic model [28, 44–47]. To date these segmentations have been performed using custom software that cannot be moved from one lab to another. Also, previously studied 3D printing software were not FDA cleared for an intended use of 3D printing cardiac anatomic models [48]. The potential nearly 21x reduction of segmentation time from roughly 72 to roughly 3.5 minutes would be a dramatic workflow improvement, while acknowledging that 3D printing itself had a mean time of over 18 hours. Even if the segmentation time was reduced from 70 to 10 minutes, the one-hour gain would prove highly valuable, since it reduces human effort, and given the fact that personnel expenses constitute the largest portion of 3D printed anatomic models [42].

The Bland-Altman analysis showed a negative bias in the data comparing DLS volumes to the SS volumes. The primary reason for this is that the SS techniques used a wider threshold band for segmentation and underwent filling of holes in a subsequent operation (Fig. 7). The significant difference between the DLS and SS volume data is also attributed to this. The DSC is one of the most popular metrics to evaluate segmentation accuracy and provides the volume overlap; the ASSD provides an “averaged” distance between surfaces of the models being compared, and the HD is the longest distance from a point in one set to the closest point in the second set [41]. In this study the DSC was high with an average of 0.96 across the 39 models, which is higher than the DSC of 0.91–0.94 reported in other studies [26–29]. The average ASSD was 0.61mm which is less than 1mm further demonstrating a good match between the model surfaces. The average HD was 8.8 mm due to the inclusion of a coronary vessel in the deep learning segmentation in some of the patients, but it is similar compared to the HD of 6.00-10.25mm in other reports [27, 29]. The average relative error (5.74%) was lower than that (7.5%) in a published study [27]. These results were obtained for 4 different appendage morphologies and a range of LA volumes across the 39 models, thus demonstrating the robust performance of the DLS algorithm.

The IC reported increased confidence and the perception of higher safety when the 3D printed anatomic model was used for procedural planning. Successful deployment of a LAA exclusion device requires a detailed intuition of the patient-specific geometry, and this may be challenging using 3D visualization alone [15, 18, 49]. The flexible model allowed the IC to size the occluder device as well as rehearse the implantation depth and orientation. The flexible 3D printed model allowed the orifice to become symmetric upon device implantation during simulation [50]. The translucency of the flexible material allowed visualization of the occlusion and delivery device from the outside the LAA. Although not detailed as part of this research, the model also proved to be a useful tool for communicating the procedure with the patient, trainees, and industry representatives. The average number of devices attempted across the 34 patients who received a device was 1.3, which was similar to the 1.245 devices attempted in a published study with 53 patients using 3D printing, and lower than the 1.6 devices attempted in a published study with 72 retrospective and 32 prospective patients using 3D printing [51, 52]. No patient in this cohort had a long-term complication.

The present study has inherent limitations from its observational nature and single-center design and the relatively small sample size. Meticulous quality control would still be needed should this method eventually supplant more manual segmentation strategies. Second, we manually trimmed the distal segments of the PVs on both the standard and deep learning segmented volumes, which could have had a minor effect on the comparison data. Finally, flexible vs rigidly 3D printed anatomic models of the LAA were not compared, although both were available for the a certain number of patients, since the IC preferred to utilize only flexible models.

Flexible 3D printed anatomic models of the LAA are clinically useful in simulating the device size for occlusion. There were no major complications in the 34 patients that underwent procedure using our 3D printed models for pre-procedural planning and simulation. Our DLS demonstrated excellent performance against the SS. DLS could serve as a powerful starting point for clinical 3D printing tasks.

Abbreviation	Full Form
AF	Atrial Fibrillation
LAA	Left Atrial Appendage
3D	Three Dimensional
AI	Artificial Intelligence
LA	Left Atrium
PV	Pulmonic Vein
CT	Computer Tomography
ECG	Electrocardiogram
DICOM	Digital Imaging and Communications in Medicine
2D	Two Dimensional
DLS	Deep Learning Segmentation
SS	Standard Segmentation
DSC	Dice Similarity Coefficient
ASSD	Average Symmetric Surface Distance
HD	Hausdorff Distance
STL	Stereolithography
TEE	Transesophageal Echocardiography
CAD	Computer Aided Design
CI	Cardiovascular Imager
IC	Interventional Cardiologist

Ethics approval and consent to participate

The study was approved by the University of Cincinnati College of Medicine Institutional Review Board.

Consent for publication

Not applicable.

Availability of data and materials

De-identified data associated with the anatomic models can be shared upon request.

Competing interests

PS and PL were employees of Materialise at the time the deep learning portion of the study was conducted. Once the deep learning output was generated by Materialise, the subsequent analyses were all conducted by PR, LLC, and AAG at the University of Cincinnati and University Hospital Zurich.

Funding

None.

Authors' contributions

FJR and PR planned the study. PR, MB, and SK performed the image segmentation in consultation with SB and AC. Image segmentation was validated by FJR. PR, SK, SC, KEW, and MAF 3D printed the anatomic models. SF and IYL electronically reviewed the patient charts and consulted with SB and AC for patient follow-up. PS and PL performed the deep learning segmentation with LLC, AAG, and PR, all 3 of whom contributed to the analyses. All authors provided textual contributions and approved the final manuscript.

Acknowledgements

Not applicable.

Kornej J, Börschel CS, Börschel CS, Benjamin EJ, Benjamin EJ, Schnabel RB, Schnabel RB. Epidemiology of Atrial Fibrillation in the 21st Century: Novel Methods and New Insights, Circ Res (2020) 4–20. doi:10.1161/CIRCRESAHA.120.316340.
Tzikas A, Shakir S, Gafoor S, Omran H, Berti S, Santoro G, Kefer J, Landmesser U, Nielsen-Kudsk JE, Cruz-Gonzalez I, Sievert H, Tichelbäcker T, Kanagaratnam P, Nietlispach F, Aminian A, Kasch F, Freixa X, Danna P, Rezzaghi M, Vermeersch P, Stock F, Stolcova M, Costa M, Ibrahim R, Schillinger W, Meier B, Park JW. Left atrial appendage occlusion for stroke prevention in atrial fibrillation: Multicentre experience with the AMPLATZER Cardiac Plug, EuroIntervention. 11 (2016) 1170–1179. doi:10.4244/EIJY15M01_06.
Korsholm K, Berti S, Iriart X, Saw J, Wang DD, Cochet H, Chow D, Clemente A, De Backer O, Møller J, Jensen JE. Nielsen-Kudsk, Expert Recommendations on Cardiac Computed Tomography for Planning Transcatheter Left Atrial Appendage Occlusion, JACC Cardiovasc. Interv. 2020;13:277–92. doi:10.1016/j.jcin.2019.08.054.
So CY, Kang G, Villablanca PA, Ignatius A, Asghar S, Dhillon D, Lee JC, Khan A, Singh G, Frisoli TM, O’neill BP, Eng MH, Song T, Pantelic M, O’neill WW, Wang DD. Additive value of preprocedural computed tomography planning versus stand-alone transesophageal echocardiogram guidance to left atrial appendage occlusion: Comparison of real-world practice. J Am Heart Assoc. 2021;10:1–8. doi:10.1161/JAHA.120.020615.
Giannopoulos AA, Mitsouras D, Yoo SJ, Liu PP, Chatzizisis YS, Rybicki FJ. Applications of 3D printing in cardiovascular diseases. Nat Rev Cardiol. 2016;13:701–18. doi:10.1038/nrcardio.2016.170.
Iriart X, Ciobotaru V, Martin C, Cochet H, Jalal Z, Thambo JB, Quessard A. Role of cardiac imaging and three-dimensional printing in percutaneous appendage closure. Arch Cardiovasc Dis. 2018;111:411–20. doi:10.1016/j.acvd.2018.04.005.
Mitsouras D, Liacouras P, Imanzadeh A, Giannopoulos AA, Cai T, Kumamaru KK, George E, Wake N, Caterson EJ, Pomahac B, Ho VB, Grant GT, Rybicki FJ. Medical 3D printing for the radiologist. Radiographics. 2015;35:1965–88. doi:10.1148/rg.2015140320.
Chepelev L, Wake N, Ryan J, Althobaity W, Gupta A, Arribas E, Santiago L, Ballard DH, Wang KC, Weadock W, Ionita CN, Mitsouras D, Morris J, Matsumoto J, Christensen A, Liacouras P, Rybicki FJ, Sheikh A. Radiological Society of North America (RSNA) 3D printing Special Interest Group (SIG): guidelines for medical 3D printing and appropriateness for clinical scenarios, 3D Print Med 4 (2018). doi:10.1186/s41205-018-0030-y.
Ravi P, Chepelev LL, Stichweh GV, Jones BS, Rybicki FJ. Medical 3D Printing Dimensional Accuracy for Multi-pathological Anatomical Models 3D Printed Using Material Extrusion. J Digit Imaging. 2022;35:613–22. doi:10.1007/s10278-022-00614-x.
George E, Liacouras P, Rybicki FJ, Mitsouras D. Measuring and establishing the accuracy and reproducibility of 3D printed medical models. Radiographics. 2017;37:1424–50. doi:10.1148/rg.2017160165.
Bastawrous S, Wu L, Strzelecki B, Levin DB, Li JS, Coburn J, Ripley B. Establishing quality and safety in hospital-based 3d printing programs: Patient-first approach. Radiographics. 2021;41:1208–29. doi:10.1148/rg.2021200175.
Cai T, Rybicki FJ, Giannopoulos AA, Schultz K, Kumamaru KK, Liacouras P, Demehri S, Shu Small KM, Mitsouras D. The residual STL volume as a metric to evaluate accuracy and reproducibility of anatomic models for 3D printing: application in the validation of 3D-printable models of maxillofacial bone from reduced radiation dose CT images, 3D Print. Med. 2015;1:1–9. doi:10.1186/s41205-015-0003-3.
Giannopoulos AA, Steigner ML, George E, Barile M, Hunsaker AR, Rybicki FJ, Mitsouras D, Cardiothoracic Applications of 3D Printing., 2016. doi:10.1097/RTI.0000000000000217.
Li H, Qingyao, Bingshen M, Shu, Lizhong X, Wang Z, Song. Application of 3D printing technology to left atrial appendage occlusion. Int J Cardiol. 2017;231:258–63. doi:10.1016/j.ijcard.2017.01.031.
DeCampos D, Teixeira R, Saleiro C, Oliveira-Santos M, Paiva L, Costa M, Botelho A, Gonçalves L. 3D printing for left atrial appendage closure: A meta-analysis and systematic review. Int J Cardiol. 2022;356:38–43. doi:10.1016/j.ijcard.2022.03.042.
Goitein O, Fink N, Guetta V, Beinart R, Brodov Y, Konen E, Goitein D, Di Segni E, Grupper A, Glikson M. Printed MDCT 3D models for prediction of left atrial appendage (LAA) occluder device size: A feasibility study, EuroIntervention. 13 (2017) e1076–e1079. doi:10.4244/EIJ-D-16-00921.
Liu P, Liu R, Zhang Y, Liu Y, Tang X, Cheng Y. The value of 3D printing models of left atrial appendage using real-time 3D transesophageal echocardiographic data in left atrial appendage occlusion: Applications toward an era of truly personalized medicine. Cardiol. 2016;135:255–61. doi:10.1159/000447444.
Hong D, Moon S, Cho Y, Oh I-Y, Chun EJ, Kim N. Rehearsal simulation to determine the size of device for left atrial appendage occlusion using patient-specific 3D-printed phantoms. Sci Rep. 2022;12:1–11. doi:10.1038/s41598-022-11967-2.
Bieliauskas G, Otton J, Chow DHF, Sawaya FJ, Kofoed KF, Søndergaard L, De Backer O. Use of 3-Dimensional Models to Optimize Pre-Procedural Planning of Percutaneous Left Atrial Appendage Closure, JACC Cardiovasc. Interv. 2017;10:1067–70. doi:10.1016/j.jcin.2017.02.027.
Kaafarani M, Saw J, Daniels M, Song T, Rollet M, Kesinovic S, Lamorgese T, Kubiak K, Qi Z, Pantelic M, O’Neill W, Wang DD. Role of CT imaging in left atrial appendage occlusion for the WATCHMAN™ device. Cardiovasc Diagn Ther. 2020;10:45–58. doi:10.21037/cdt.2019.12.01.
Vivoli G, Gasparotti E, Rezzaghi M, Cerone E, Mariani M, Landini L, Berti S, Positano V, Celi S, Simultaneous Functional and Morphological Assessment of Left Atrial Appendage by 3D Virtual Models, J. Healthc. Eng. 2019 (2019). doi:10.1155/2019/7095845.
Litjens G, Ciompi F, Wolterink JM, de Vos BD, Leiner T, Teuwen J, Išgum I. State-of-the-Art Deep Learning in Cardiovascular Image Analysis. JACC Cardiovasc Imaging. 2019;12:1549–65. doi:10.1016/j.jcmg.2019.06.009.
Baskaran L, Al’Aref SJ, Maliakal G, Lee BC, Xu Z, Choi JW, Lee SE, Sung JM, Lin FY, Dunham S, Mosadegh B, Kim YJ, Gottlieb I, Lee BK, Chun EJ, Cademartiri F, Maffei E, Marques H, Shin S, Choi JH, Chinnaiyan K, Hadamitzky M, Conte E, Andreini D, Pontone G, Budoff MJ, Leipsic JA, Raff GL, Virmani R, Samady H, Stone PH, Berman DS, Narula J, Bax JJ, Chang HJ, Min JK, Shaw LJ. Automatic segmentation of multiple cardiovascular structures from cardiac computed tomography angiography images using deep learning. PLoS One. 2020;15:1–13. doi:10.1371/journal.pone.0232573.
Jaskari J, Sahlsten J, Järnstedt J, Mehtonen H, Karhu K, Sundqvist O, Hietanen A, Varjonen V, Mattila V, Kaski K. Deep Learning Method for Mandibular Canal Segmentation in Dental Cone Beam Computed Tomography Volumes. Sci Rep. 2020;10:1–8. doi:10.1038/s41598-020-62321-3.
Leventic H, Bencevic M, Babin D, Habijan M, A Survey of Left Atrial Appendage Segmentation and Analysis in 3D and 4D Medical Images, 28th Int. Conf. Syst. Signals Image Process. (2021) 1–12.
Qiao M, Wang Y, Berendsen FF, van der Geest RJ, Tao Q. Fully automated segmentation of the left atrium, pulmonary veins, and left atrial appendage from magnetic resonance angiography by joint-atlas-optimization. Med Phys. 2019;46:2074–84. doi:10.1002/mp.13475.
Xiong Z, Fedorov VV, Fu X, Cheng E, Macleod R, Zhao J. Fully Automatic Left Atrium Segmentation From Late Gadolinium Enhanced Magnetic Resonance Imaging Using a Dual Fully Convolutional Neural Network. IEEE Trans Med Imaging. 2019;38:515–24. doi:10.1109/TMI.2018.2866845.
Jia S, Despinasse A, Wang Z, Delingette H, Pennec X, Jaïs P, Cochet H, Sermesant M, Automatically Segmenting the Left Atrium from Cardiac Images Using Successive 3D U-Nets and a Contour Loss, Lect. Notes Comput. Sci. (Including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics). 11395 LNCS (2019) 221–229. doi:10.1007/978-3-030-12029-0_24.
Zhang X, Martin DG, Noga M, Punithakumar K, Fully Automated Left Atrial Segmentation from MR Image Sequences Using Deep Convolutional Neural Network and Unscented Kalman Filter, Proc. – 2018 IEEE Int. Conf. Bioinforma. Biomed. BIBM 2018. (2019) 2316–2323. doi:10.1109/BIBM.2018.8621570.
Rybicki FJ, Otero HJ, Steigner ML, Vorobiof G, Nallamshetty L, Mitsouras D, Ersoy H, Mather RT, Judy PF, Cai T, Coyner K, Schultz K, Whitmore AG. M.F. Di Carli, Initial evaluation of coronary images from 320-detector row computed tomography. Int J Cardiovasc Imaging. 2008;24:535–46. doi:10.1007/s10554-008-9308-2.
Steigner ML, Mitsouras D, Whitmore AG, Otero HJ, Wang C, Buckley O, Levit NA, Hussain AZ, Cai T, Mather RT, Smedby Ö, Dicarli MF, Rybicki FJ. Iodinated contrast opacification gradients in normal coronary arteries imaged with prospectively ECG-gated single heart beat 320-detector row computed tomography. Circ Cardiovasc Imaging. 2010;3:179–86. doi:10.1161/CIRCIMAGING.109.854307.
Kumamaru KK, Hoppel BE, Mather RT, Rybicki FJ. CT angiography: current technology and clinical use. Radiol Clin North Am. 2010;48:213–35. doi:10.1016/j.rcl.2010.02.006.
Pour-Ghaz I, Heckle MR, Maturana M, Seitz MP, Zare P, Khouzam RN, Kabra R. Percutaneous Left Atrial Appendage Closure: Review of Anatomy, Imaging, and Outcomes, Curr. Treat. Options Cardiovasc Med. 2022;24:41–59. doi:10.1007/s11936-022-00958-1.
Lee AP, Vida VL, Cardiovascular 3D Printing, 2021. doi:10.1007/978-981-15-6957-9.
Chepelev L, Souza C, Althobaity W, Miguel O, Krishna S, Akyuz E, Hodgdon T, Torres C, Wake N, Alexander A, George E, Tang A, Liacouras P, Matsumoto J, Morris J, Christensen A, Mitsouras D, Rybicki F, Sheikh A. Preoperative planning and tracheal stent design in thoracic surgery: a primer for the 2017 Radiological Society of North America (RSNA) hands-on course in 3D printing, 3D Print. Med 3 (2017). doi:10.1186/s41205-017-0022-3.
Chepelev L, Hodgdon T, Gupta A, Wang A, Torres C, Krishna S, Akyuz E, Mitsouras D, Sheikh A. Medical 3D printing for vascular interventions and surgical oncology: a primer for the 2016 radiological society of North America (RSNA) hands-on course in 3D printing, 3D Print. Med 2 (2016). doi:10.1186/s41205-016-0008-6.
Giannopoulos AA, Chepelev L, Sheikh A, Wang A, Dang W, Akyuz E, Hong C, Wake N, Pietila T, Dydynski PB, Mitsouras D, Rybicki FJ. 3D printed ventricular septal defect patch: a primer for the 2015 Radiological Society of North America (RSNA) hands-on course in 3D printing, 3D Print. Med. 2015;1:1–20. doi:10.1186/s41205-015-0002-4.
Ravi P, Chepelev L, Lawera N, Haque KMA, Chen VCP, Ali A, Rybicki FJ. A systematic evaluation of medical 3D printing accuracy of multi-pathological anatomical models for surgical planning manufactured in elastic and rigid material using desktop inverted vat photopolymerization. Med Phys. 2021;48:3223–33. doi:10.1002/mp.14850.
Ren S, He K, Girshick R, Sun J, Faster R-CNN: Towards real-time object detection with region proposal networks, in: NeurIPS Proc., 2015: pp. 1–9. doi:10.4324/9780080519340-12.
Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. In: Navab N, Hornegger J, Wells WM, Frangi AF, editors. Med. Image Comput. Comput. Interv. -- MICCAI 2015. Cham: Springer International Publishing; 2015. pp. 234–41.
Heimann T, Van Ginneken B, Styner MA, Arzhaeva Y, Aurich V, Bauer C, Beck A, Becker C, Beichel R, Bekes G, Bello F, Binnig G, Bischof H, Bornik A, Cashman PMM, Chi Y, Córdova A, Dawant BM, Fidrich M, Furst JD, Furukawa D, Grenacher L, Hornegger J, Kainmüller D, Kitney RI, Kobatake H, Lamecker H, Lange T, Lee J, Lennon B, Li R, Li S, Meinzer HP, Németh G, Raicu DS, Rau AM, Van Rikxoort EM, Rousson M, Ruskó L, Saddi KA, Schmidt G, Seghers D, Shimizu A, Slagmolen P, Sorantin E, Soza G, Susomboon R, Waite JM, Wimmer A, Wolf I. Comparison and evaluation of methods for liver segmentation from CT datasets. IEEE Trans Med Imaging. 2009;28:1251–65. doi:10.1109/TMI.2009.2013851.
Ravi P, Burch MB, Farahani S, Chepelev LL, Yang D, Ali A, Joyce JR, Lawera N, Stringer J, Morris JM, Ballard DH, Wang KC, Kondor S, Rybicki FJ, U of CCS Participants, Utility and costs during the initial one-year of 3D printing in an academic hospital, J. Am. Coll. Radiol. (2022). doi:10.1016/j.jacr.2022.07.001.
Fan Y, Wong RHL, Lee AP-W. Three-dimensional printing in structural heart disease and intervention. Ann Transl Med. 2019;7:579–9. doi:10.21037/atm.2019.09.73.
Leventić H, Babin D, Velicki L, Devos D, Galić I, Zlokolica V, Romić K, Pižurica A. Left atrial appendage segmentation from 3D CCTA images for occluder placement procedure. Comput Biol Med. 2019;104:163–74. doi:10.1016/j.compbiomed.2018.11.006.
Wang DD, Qian Z, Vukicevic M, Engelhardt S, Kheradvar A, Zhang C, Little SH, Verjans J, Comaniciu D, O’Neill WW. M.A. Vannan, 3D Printing, Computational Modeling, and Artificial Intelligence for Structural Heart Disease. JACC Cardiovasc Imaging. 2021;14:41–60. doi:10.1016/j.jcmg.2019.12.022.
Liu H, Feng J, Feng Z, Lu J, Zhou J. Left atrium segmentation in CT volumes with fully convolutional networks, Lect Notes Comput Sci (Including Subser Lect Notes Artif Intell Lect Notes Bioinformatics) 10553 LNCS (2017) 39–46. doi:10.1007/978-3-319-67558-9_5.
Vesal S, Ravikumar N, Maier A. Dilated Convolutions in Neural Networks for Left Atrial Segmentation in 3D Gadolinium Enhanced-MRI, Lect Notes Comput Sci (Including Subser Lect Notes Artif Intell Lect Notes Bioinformatics) 11395 LNCS (2019) 319–28. doi:10.1007/978-3-030-12029-0_35.
Mitsouras D, Liacouras PC, Wake N, Rybicki FJ. Radiographics update: Medical 3d printing for the radiologist. Radiographics. 2020;40:E21–3. doi:10.1148/rg.2020190217.
Harb SC, Rodriguez LL, Vukicevic M, Kapadia SR, Little SH. Three-Dimensional Printing Applications in Percutaneous Structural Heart Interventions. Circ Cardiovasc Imaging. 2019;12:1–11. doi:10.1161/CIRCIMAGING.119.009014.
Little SH, Saric M. Three-Dimensional Printing and the Auricle: Predicting Future Events? J Am Soc Echocardiogr. 2019;32:720–1. doi:10.1016/j.echo.2019.04.005.
Wang DD, Eng M, Kupsky D, Myers E, Forbes M, Rahman M, Zaidan M, Parikh S, Wyman J, Pantelic M, Song T, Nadig J, Karabon P, Greenbaum A, O’Neill W. Application of 3-Dimensional Computed Tomographic Image Guidance to WATCHMAN Implantation and Impact on Early Operator Learning Curve: Single-Center Experience, JACC Cardiovasc. Interv. 2016;9:2329–40. doi:10.1016/j.jcin.2016.07.038.
Fan Y, Yang F, Cheung GSH, Chan AKY, Wang DD, Lam YY, Chow MCK, Leong MCW, Kam KKH, So KCY, Tse G, Qiao Z, He B, Kwok KW, Lee APW. Device Sizing Guided by Echocardiography-Based Three-Dimensional Printing Is Associated with Superior Outcome after Percutaneous Left Atrial Appendage Occlusion. J Am Soc Echocardiogr. 2019;32:708–19.e1. doi:10.1016/j.echo.2019.02.003.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

3D printed flexible anatomical models for left atrial appendage closure planning and comparison of deep learning against radiologist image segmentation

Status:

Version 1

Abstract

Figures

Background

Methods

Part 1: Clinical sub-study

Patients

Clinical Cardiac CT

3D Printing Standard Segmentation (SS)

Computer Aided Design and 3D Printing

Watchman Device Procedural Planning and Deployment

Patient Clinical Follow Up

Part 2. Study of AI based segmentation for LAA occlusion planning

Deep Learning Segmentation (DLS)

Test Metrics for the AI Segmentation

Statistical Analysis

Results

Patients

Clinical Cardiac CT

3D Printing Standard Segmentation (SS)

Computer Aided Design and 3D Printing of Models

Watchman Device Procedural Planning and Deployment

Patient Clinical Follow Up

Deep Learning Segmentation (DLS)

Test Metrics

Discussion

Conclusions

Abbreviations

Declarations

References

Additional Declarations

Status:

Version 1