Abstract
Background Computed tomography (CT)-derived pectoralis muscle area (PMA) measurements are prognostic in people with or at-risk of COPD, but fully automated PMA extraction has yet to be developed. Our objective was to develop and validate a PMA extraction pipeline that can automatically: 1) identify the aortic arch slice; and 2) perform pectoralis segmentation at that slice.
Methods CT images from the Canadian Cohort of Obstructive Lung Disease (CanCOLD) study were used for pipeline development. Aorta atlases were used to automatically identify the slice containing the aortic arch by group-based registration. A deep learning model was trained to segment the PMA. The pipeline was evaluated in comparison to manual segmentation. An external dataset was used to evaluate generalisability. Model performance was assessed using the Dice–Sorensen coefficient (DSC) and PMA error.
Results In total 90 participants were used for training (age 67.0±9.9 years; forced expiratory volume in 1 s (FEV1) 93±21% predicted; FEV1/forced vital capacity (FVC) 0.69±0.10; 47 men), and 32 for external testing (age 68.6±7.4 years; FEV1 65±17% predicted; FEV1/FVC 0.50±0.09; 16 men). Compared with manual segmentation, the deep learning model achieved a DSC of 0.94±0.02, 0.94±0.01 and 0.90±0.04 on the true aortic arch slice in the train, validation and external test sets, respectively. Automated aortic arch slice detection obtained distance errors of 1.2±1.3 mm and 1.6±1.5 mm on the train and test data, respectively. Fully automated PMA measurements were not different from manual segmentation (p>0.05). PMA measurements were different between people with and without COPD (p=0.01) and correlated with FEV1 % predicted (p<0.05).
Conclusion A fully automated CT PMA extraction pipeline was developed and validated for use in research and clinical practice.
Tweetable abstract
A fully automated pipeline for the extraction of pectoralis muscle area measurements from low-dose chest CT was developed that was highly accurate and externally validated https://bit.ly/3R1XgD5
Introduction
COPD is a lung condition characterised by airflow limitation. Recent studies show extrapulmonary comorbidities, such as involuntary loss of muscle mass (sarcopenia), begin to manifest in at-risk individuals [1] and have been linked with all-cause mortality [2]. In people diagnosed with COPD, low fat-free muscle mass (FFM) measurements are associated with impaired exercise tolerance [3] and increase the risk of mortality, irrespective of pulmonary function [4]. However, FFM measurements are not typically acquired as part of COPD clinical care.
Computed tomography (CT) images are routinely collected in at-risk individuals who smoke as part of lung cancer screening trials [5] and in people with pulmonary abnormalities [6]. The pectoralis muscle area (PMA) is a measure that can be extracted from chest CT, and has been correlated with FFM [7] and associated with prognosis in a variety of populations, such as at-risk for [8] and with COPD [9], and interstitial lung disease (ILD) [10]. The established methodology to extract PMA measurements is: 1) manually identify the slice containing the aortic arch; and 2) perform semi-automated segmentation of the pectoralis at that slice [7–10]. Although the PMA has been manually extracted in large cohorts such as COPD Genetic Epidemiology study (COPDGene) [11] and ECLIPSE (Evaluation of COPD Longitudinally to Identify Predictive Surrogate End-points [12]) [7–9, 13], and is a target biomarker for unexplored cohorts, such as CanCOLD (Canadian Cohort of Obstructive Lung Disease [14]), it is impractical to perform segmentation in routine clinical care as the process is time-consuming and introduces intra/interobserver variability. Therefore, a fully automated method to extract the PMA is required for use in routine clinical care.
Convolutional neural networks (CNNs) are deep learning-based tools that are capable of generating predictions using image inputs. Specifically, the U-Net model is a commonly used CNN architecture that specialises in image segmentation and produces highly accurate segmentation results as compared to manual segmentations [15, 16]. Multiple studies have trained such models to automatically extract body composition measurements in abdominal CT with high accuracy [17, 18]. However, few studies have developed an automated method to quantify the PMA from chest scans [19, 20], and no studies have performed both automatic identification of the aortic arch and pectoralis segmentation.
Therefore, our primary objective was to develop a fully automated pipeline using the CanCOLD study of people with or without primarily mild COPD and test its performance in an external group of research study participants with COPD [21]. To accomplish this objective, we aimed to: 1) create an algorithm to automatically identify the aortic arch slice; and 2) train a deep learning model to automatically segment the pectoralis muscle in the identified slice. Our secondary objective was to investigate the discriminative and concurrent validity of the automated PMA measurements by quantifying differences between subgroups and correlations to pulmonary function test variables respectively.
Materials and methods
Dataset
CanCOLD is a multi-centre cohort of ∼1800 participants aged ≥40 years originally sampled through random dialling across nine sites in Canada [14]. A subset of 90 participants (10 per site) from the baseline visit were randomly selected for the development of the automated pipeline; the participants maintained an equal ratio of males/females, people with/without COPD, and smokers/never-smokers. Informed written consent from the CanCOLD participants and approval by institutional review boards was obtained at each site. All demographics, pulmonary function and chest CT scans were retrospectively collected at baseline. CT analysis of all CanCOLD participants has been previously reported [22–29]; however, the prior studies dealt with development of lung CT measurements, whereas in this manuscript we report development of muscle CT measurements. A group of people with COPD studied at the University of Alberta (UofA) (iNO-NCT03679312) were used to externally test the pipeline [21]. The UofA cohort derives from a single-centre interventional study of people 18–85 years old that focused on examining the effects of inhaled nitric oxide on exercise capacity in COPD patients. Participants were excluded if there was no evidence of left and/or right pectoralis minor and major muscle at the level of the aortic arch.
Pulmonary function measurements
Spirometry was performed according to American Thoracic Society (ATS) guidelines [30–32] for measurement of the postbronchodilator forced expiratory volume in 1 s (FEV1) and forced vital capacity (FVC) [13]. The residual volume/total lung capacity (RV/TLC) ratio was measured using whole-body plethysmography.
Image acquisition
Chest CT images were acquired at full inspiration using various systems across all CanCOLD sites with the following parameters: 120 kVp, 1.0- or 1.25-mm slice thickness, and 0.52–0.90 mm2 pixel spacing. The external test set from the UofA was acquired using a similar protocol: 120 kVp, 0.75 mm slice thickness, and 0.55–0.75 mm2 pixel spacing. A standard or soft tissue kernel was used for image reconstruction. All key acquisition parameters from both training and test images are listed in supplementary table E1.
Ground truth pectoralis muscle segmentation
Two observers (A.R. Jenkins, a postdoctoral student with 1 year of experience, and N. van Noord, a PhD student with 1 year of experience) performed ground truth pectoralis muscle segmentations (45 per observer) using the Chest Imaging Platform extension in 3D Slicer (https://cip.bwh.harvard.edu/index.html) [7]. Briefly, the observers manually identified the axial slice containing the aortic arch. The pectoralis muscle was then segmented five slices below to five slices above the aortic arch using a manual brush tool, creating 11 slices/participant (990 slices total) for deep learning model training. Both observers then repeated segmentations from a random subset of 15 participants to quantify intra- and interobserver variability. In the external dataset, each observer segmented the pectoralis muscle at the aortic arch slice from 16 CT scans (32 slices total) for testing.
Aortic arch slice determination
Aorta segmentation
All proceeding pipeline development was performed in MATLAB R2021a (v. 9.10). First, all CT volumes were resized to a resolution of 1 mm3 using bicubic interpolation. The lungs were then automatically segmented using thresholding and volumes cropped to localise the top of the heart. In the remaining slices, contrast was rescaled and images were smoothed of high-frequency artefacts using a median filter. Ray casting was used to extract the heart cavity between the lungs. The heart muscle was then automatically segmented using K-means segmentation (n=4 centres) and cleaned of peripheral blood vessels, creating 90 aorta atlases. A visual representation of the aorta segmentation process is shown in supplementary figure E1. The observers also manually delineated the aortic arch from the aorta segmentations to create 90 aortic arch atlases using the “VolumeSegmenter” application, as visualised in supplementary figure E2. A more detailed description is provided in the online supplementary material.
Aortic arch detection
The aortic arch was detected in the aorta segmentations using an image-registration-based approach. A candidate aorta segmentation was compared to all aorta atlases using a cross-correlation of their voxel counts per slice, as shown in supplementary figure E3. The five most correlated atlas aortas were registered to the candidate using an affine algorithm. Resulting warp matrices were used to register the corresponding aortic arch atlases to the candidate. The top-most slice from the consensus volume, created using a majority vote of the warped atlases, was labelled as the top of the aortic arch.
Deep learning model
Image preprocessing
The input images and ground truth segmentations were resized to a matrix dimension of 256×256 for model training using bicubic and nearest neighbour interpolation, respectively. Any pixels below −190 HU and above 150 HU were set to those thresholds, and the result was rescaled to [0,1].
Network architecture
A diagram of pipeline development during training and evaluation is shown in figure 1. The 2D U-Net architecture was selected for the deep learning model [15]. The convolutional path for the model contained four down-sampling and four up-sampling blocks. A 1×1 kernel convolution followed by sigmoid activation was used to create pectoralis muscle probability maps. At the start of each epoch, data augmentation was applied to mimic natural variations in pectoral anatomy. 10-fold cross-validation was implemented during training, with convergence occurring once validation performance failed to improve compared to the previous epoch. A more detailed listing of the hyperparameters is available in supplementary table E2. The model was trained using an NVIDIA RTX A2000 12 GB GPU.
Deep learning model training and evaluation data flow diagram. a) Model training: 90 CanCOLD participants had the aortic arch slice labelled by human observers, and ground truth segmentations were performed between the five slices before and after the aortic arch slice (creating a total of 990 slices). 10-fold cross-validation was implemented during model training; b) Model evaluation: model performance was then strictly evaluated on the aortic arch slice labelled by the human observers on the 90 CanCOLD and 32 UofA participants. CanCOLD: Canadian Cohort Obstructive Lung Disease study; UofA: University of Alberta; PMA: pectoralis muscle area.
Image postprocessing
A diagram of the complete automated pipeline is shown in figure 2. The resulting probability maps were resized back to their 1 mm3 matrix size. A threshold of 0.4 was used to binarise the maps into pectoralis segmentations. Any objects <0.5 cm2 or one-third of the mean area of all components was treated as noise and removed. As a final measure, the resulting segmentations were morphologically closed (structure=square, radius=2 mm). The PMA was defined as the total area of all remaining structures. The code used for the development of the complete pipeline is available upon request.
The complete fully automated pectoralis segmentation pipeline: a) Aortic arch slice detection: first, the aorta is automatically segmented from the contrast rescaled chest computed tomography (CT). Next, the five most matching atlas arches are registered to the segmented aorta. Majority voting between the five registered atlases is used to quantify the top of the aortic arch. b) Pectoralis muscle area (PMA) extraction: the CT slice identified in the earlier step is then fed into the trained deep learning model to obtain the pectoralis muscle segmentation and subsequent pectoralis muscle area measurement.
Statistical analysis
Statistical analyses were performed using MATLAB R2021a (v. 9.10). Group differences were evaluated using the Mann–Whitney U-test or the Kruskal–Wallis test (Tukey correction). Aortic arch detection error was assessed through a difference in millimetres between the slice predicted by the pipeline and the human observers. The Sorensen–Dice Coefficient (DSC) and Jaccard index are two metrics that are typically used to evaluate predicted segmentation overlap with manual segmentation as the ground truth (0: no overlap; 1: complete overlap). Therefore, intra- and interobserver variability was evaluated using the DSC, Jaccard index and absolute (cm2)/relative (%) PMA difference. Deep learning performance at the slice labelled by the observers (PMAModel) and predicted by the pipeline (PMAPipeline) were compared to the ground truth (PMATrue) using the same metrics and a Bland–Altman analysis. Differences in the PMAPipeline measurements for various subgroups (male versus female, body mass index (BMI) >25 kg·m−2 versus BMI ≤25 kg·m−2, COPD versus no COPD) in the entire dataset (n=122) were quantified using the Mann–Whitney U-test or an ANCOVA to evaluate discriminant validity. Correlations between PMATrue and PMAPipeline with BMI, FEV1 % predicted, FEV1/FVC and RV/TLC were assessed using the Spearman's rank correlation coefficient to evaluate convergent validity. Statistical significance was evaluated at a type I error of 5%.
Results
Participant characteristics
A single participant was excluded from the UofA test cohort due to the absence of a pectoralis muscle, and no participants in either cohort had visible breast implants at the level of the aortic arch. The demographics and pulmonary function measurements for the training and testing cohorts are reported in table 1. Age, sex, race and BMI were similar between training and test sets (p>0.05). However, compared to the CanCOLD training cohort, the UofA test cohort had a greater history of smoke exposure, with evidence of airflow obstruction, and pulmonary gas trapping (p<0.05). The demographics and pulmonary function for participants stratified by Global Initiative for Obstructive Lung Disease (GOLD) stage [33] in the combined cohorts are shown in supplementary table E3.
Demographics and pulmonary function for the participants used during model training (CanCOLD) and external testing (UofA)
Inter- and intraobserver variability
Supplementary table E4 shows the intra- and interobserver variability for the manual PMA measurements. The mean intra- and interobserver absolute slice difference between the aortic arch labels were 0.2±0.6 mm and 0.5±0.6 mm, respectively. Intra- and intersegmentation overlap was 0.96±0.02 and 0.95±0.02 as measured by the DSC. The intra- and interabsolute (and relative) PMA error was 0.88±0.57 cm2 (2.4±1.5%) and 1.50±0.97 cm2 (4.3±2.5%), respectively.
Deep learning segmentation model performance
During training
All deep learning models trained using the data successfully reached the conversion criteria. The time for training the models for 10-fold cross-validation was 19.4±0.4 min per fold, with a total training time of 193.7 min. Convergence during training occurred at 6.1±1.3 epochs. Lastly, the final model mean DSC was 0.94±0.01 for both the training (n=891 slices) and validation (n=99 slices) sets across all 10-folds. The best performing model over the training and validation folds was chosen to proceed with final pipeline creation.
After postprocessing
Figure 3 visually represents the trained model's segmentation capabilities on three representative participants with varying sex and COPD status in all three training, validation and testing sets. The deep learning model performance for PMAModel measurements was compared to PMATrue measurements on the single slice containing the aortic arch as labelled by the human observers (table 2). The model obtained an average DSC of 0.94±0.02 and 0.94±0.01 on the training and validation data, respectively. Absolute (and relative) PMA error was 1.33±1.42 cm2 (4.1±4.1%) and 0.99±0.51 cm2 (2.8±1.2%) on the training and validation data, respectively. The test data obtained slightly lower performance metrics, with a DSC of 0.90±0.04 and absolute (and relative) PMA error of 1.39±1.30 cm2 (5.5±5.7%). Comparison between PMAModel and PMATrue measurements in all participants (n=122) is shown in figure 4. The PMAModel and PMATrue measurements were similar (p=0.94), highly correlated (ρ=0.98, p<0.01) and had negligible bias (b=0.49 cm2, 95% confidence interval (CI) −3.1–4.1 cm2). Individual training and test cohort comparisons, as well as model performance stratified by sex and CanCOLD centre ID, are shown in supplementary figure E4 and tables E5 and E6. We note that the pipeline performance was comparable between the sexes and between CanCOLD centres.
Deep learning model performance for an example participant from all three training, validation and testing sets. From left to right: contrast rescaled chest computed tomography (CT) containing the aortic arch as labelled by the human observers. Ground truth pectoralis muscle segmentation. Fully automated segmentation using the developed pipeline (red), ground truth segmentation (blue) and overlap between the two (green). Top row: 66-year-old male with COPD from the training cohort; Sorensen–Dice coefficient (DSC): 0.94, Jaccard: 0.89, pectoralis muscle area (PMA) error: 1.69 cm2 (4.4%). Middle row: 67-year-old female without COPD from the validation cohort; DSC: 0.94, Jaccard: 0.89, PMA error: 1.16 cm2 (3.2%). Bottom row: 62-year-old male with COPD from the test cohort; DSC: 0.93, Jaccard: 0.88, PMA error: 2.27 cm2 (5.7%).
Deep learning model segmentation performance after postprocessing in comparison to the ground truth (PMATrue) for the single slice containing the aortic arch as labelled by the observers (PMAModel) and as predicted by the aortic arch detection algorithm in the complete automated pipeline (PMAPipeline)
Bland–Altman plots, correlations and measurement differences between the pectoralis muscle areas (PMA) of all participants (n=122). a) PMA predicted by the deep learning model (PMAModel) versus PMA segmented by the observers (PMATrue) on the human labelled aortic arch slice. b) PMA predicted by the complete automated pipeline on the detected aortic arch slice (PMAPipeline) versus PMA segmented by the observers (PMATrue) on the human labelled aortic arch slice. The Bland–Altman plots contain the coefficient of variation (CV) (top-right), and mean difference and 95% confidence intervals (1.96*sd) overlaid. The correlation plots contain the Spearman's ρ (along with the p-value) and linear regression for the two sets of data at the top left corner. The box-and-whisker plots show the median and 95% confidence interval for each measurement, along with the p-value from the measurement differences.
Complete pipeline performance
Supplementary table E7 describes the time taken to complete each step of the pipeline for the training and testing cohorts. The average absolute slice error for the aortic arch detection algorithm was 1.2±1.3 mm and 1.6±1.5 mm for training and testing sets, respectively. Performance for PMAPipeline measurements using the automated detection algorithm in comparison to PMATrue measurements is also shown in table 2. The absolute (and relative) PMA error was 1.56±1.36 cm2 (4.6±4.1%), 1.11±0.80 cm2 (3.1±2.0%) and 1.50±1.43 cm2 (6.2±6.8%) on the training, validation and test data, respectively. Comparison between PMAPipeline and PMATrue measurements for all participants (n=122) is shown in figure 4. The PMAPipeline and PMATrue measurements were similar (p=0.93), highly correlated (ρ=0.98, p<0.01), and had negligible bias (b=0.23 cm2, 95% CI −3.7–4.2 cm2). Figure E5 summarises comparisons between PMAPipeline and PMAModel measurements.
Group PMA differences: discriminant validity
Using the fully automated PMA measurements generated by the pipeline, a comparison between participants stratified by sex, BMI and COPD status is shown in figure 5. PMA measurements were significantly greater in males than females (p<0.01), people with BMI >25 kg·m−2 versus BMI ≤25 kg·m−2 (p<0.01), and people without versus with COPD (p=0.01, after adjusting for age, sex and BMI). Supplementary figure E6 shows that PMA decreased significantly across the COPD continuum: from non-COPD to GOLD II+ (p<0.05, after adjusting for age, sex and BMI). Finally, PMA was significantly lower in people with a greater amount of airflow limitation and pulmonary gas trapping (p<0.05; supplementary figure E7).
Box-and-whisker plots (median, 95% confidence intervals) of fully automated pectoralis muscle area (PMA) measurements for various subgroups in the combined CanCOLD and UofA datasets (n=122): male (n=59), 41.1±9.4 cm2 versus female (n=63), 25.5±5.6 cm2, p<0.01; BMI >25 kg·m−2 (n=85), 35.3±11.7 cm2 versus BMI ≤25 kg·m−2 (n=37), 29.3±8.0 cm2, p<0.01; no COPD (n=46), 35.5±10.9 cm2 versus COPD (n=76), 32.3±11.0 cm2, p=0.02. CanCOLD: Canadian Cohort Obstructive Lung Disease study; UofA: University of Alberta; BMI: body mass index. COPD group differences were adjusted for age, sex and BMI. *: indicates a significant difference (p<0.05).
PMA correlations: concurrent validity
Increased PMA was significantly correlated with increased BMI (PMATrue: ρ=0.18, p<0.05; PMAPipeline: ρ=0.21; p<0.05), increased FEV1/FVC (PMATrue: ρ=0.37; PMAPipeline: ρ=0.32, p<0.05) and reduced RV/TLC (PMATrue: ρ=−0.33, p<0.05; PMAPipeline: ρ=−0.31, p<0.05) (table 3). Correlations between PMA with the rest of the variables and in various subgroups of the data are shown in supplementary tables E8 and E9.
Spearman correlations between BMI and pulmonary function indices with the true pectoralis muscle area (PMA) measured by the human observers (PMATrue) and the predicted PMA measured by the automated pipeline (PMAPipeline)
Discussion
We developed a fully automated PMA extraction pipeline from chest CT images. Our deep learning model was able to segment the pectoralis muscle with high accuracy in both the validation and external test sets. Our aortic arch detection algorithm was accurate to <2 slices, and the PMAs generated by the complete pipeline on the predicted aortic arch slice was highly correlated, not significantly different and showed negligible Bland–Altman bias in comparison to the ground truths. Our fully automated PMA measurements were validated discriminately (by reporting significant differences between participants stratified by sex, BMI and COPD status) and concurrently (by reporting significant correlations with lung function indices).
The deep learning model generated pectoralis segmentation with high accuracy in comparison to the ground truth segmentation. The performance of the model was also similar to the interobserver variability of the segmentation performed by the human observers (Observers: 0.95±0.02; Model: 0.94±0.02). When combined with the significant time-savings provided by the automated segmentation in comparison to the human observers, it provides substantial evidence for the use of this pipeline to analyse large cohorts, such as CanCOLD. In addition, our model accuracy was either similar (magnetic resonance imaging-based; DSC: 0.94±0.01) [20] or slightly higher (CT-based; DSC: 0.93±0.04) [19] than other state-of-the art pectoralis segmentation networks from the literature. Unlike the previously published work, we also quantified model performance on an external never-before-seen dataset of research study participants with COPD, which is important for testing generalisability. Furthermore, we believe the slight decrease in average absolute PMA error (−1.4%) on the test cohort may be explained by the variable CT scan settings relative to the training data. This is moreover evidenced by the increased systematic bias in the Bland–Altman plots for the test cohort PMAModel measurements as compared to the training cohort (1.1 cm2 versus 0.28 cm2). Nonetheless, we believe the high model-ground truth DSC and low absolute PMA error on all datasets confirm the utility and generalisability of our deep learning model.
Our study was the first to create and report on a fully automated algorithm and complete pipeline for identifying the chest CT image slice containing the aortic arch. Similar slice localisation techniques for automated body composition measurements in abdominal CT have reported average errors of 4–5 mm using deep learning-based methods [17, 18]. In our work, the average error was <2 mm, indicating that our simpler atlas-matching approach was comparable to the state-of-the-art and sufficiently accurate to identify the aortic arch slice. Furthermore, we found that the average PMA error increased slightly when implementing the complete pipeline on the train (+0.5%) and test (+0.7%) sets. However, this did not necessarily reflect a worsening in measurement accuracy; there may have been natural minor variations in pectoral anatomy between the predicted and true aortic arch slice. In addition, PMAPipeline measurements were highly comparable to PMATrue as indicated by the Bland–Altman analysis and nearly linear positive correlation. These findings indicate that our aortic arch detection algorithm did not significantly affect the final model performance.
Finally, the results showed that PMA measurements generated by the complete pipeline were able to detect significant differences between various subgroups. In agreement with Diaz et al. [34] and McDonald et al. [7], we showed that PMA was significantly lower in: females versus males; and people with compared to without COPD. We also showed that PMA was significantly greater in people with BMI >25 kg·m−2 versus BMI ≤25 kg·m−2. Both PMATrue and PMAPipeline measurements showed similar and significant correlations with FEV1 % predicted, FEV1/FVC and RV/TLC for all participants, indicating that pipeline automation captures the same trends in the data as manual segmentation performed by human observers.
Several limitations of our work must be acknowledged. Our model was trained on a relatively small number of participants with and without COPD (n=90), which could potentially limit the generalisability for future use. However, the countermeasures undertaken by the pipeline (data augmentation, hold-out validation sets and testing on a never-before-seen external dataset) helped offset this supposed limitation. Nonetheless, future work should look to expand the pipeline using more participants with varying anatomy, including those with breast implants. We acknowledge that the pipeline's predictions may deviate from the manual observer's segmentations. However, the fully automated pipeline would reduce any measurement variability, which may allow for more subtle changes over time to be measured. The intra-/interobserver variability for a manual PMA segmentation undertaken by the same (0.9±0.6 cm2 or 2.4±1.5%) or different observer (1.5±1.0 cm2 or 4.3±2.5%), as reported in this study, would introduce measurement error, thereby reducing the potential to measure longitudinal change. For example, with manual segmentation, the reproducibility coefficient was found to be 3.5 cm2, therefore, any longitudinal change detected that is <3.5 cm2, as measured by different observers, could be attributed to measurement error [35]. Additionally, Mason et al. [13] showed that PMA longitudinal change over a 5-year span was −1.8 cm2 (−6.0%) in women and −2.8 cm2 (−6.0%) in men. This longitudinal change is within the measurement error for manual segmentation of the PMA as reported in our study. Therefore, although our pipeline slightly deviates from the ground truth, a fully automated approach reduces measurement variation error and therefore allows for smaller changes to be measured over time. We also acknowledge that other deep learning segmentation models such as YOLO [36] and DeepLab [37] should be explored. Future work should investigate the performance of each for this application. We acknowledge that there are further improvements to a fully automated PMA measurement pipeline that can potentially increase speed and accuracy. For example, developing an automated method to segment the aorta directly would remove potential errors introduced during atlas registration [38, 39]. A 3D deep learning network could bypass slice detection altogether, allowing for the entire segmentation of the pectoralis muscle volume. However, the current standard in the field is 2D segmentation at the aortic arch slice, and therefore such work would be an avenue of future research.
Limitations notwithstanding, our work is an important step forward in automated PMA extraction from chest CT scans. For example, this pipeline can be used in research settings to dichotomise people into low versus high skeletal muscle mass, to assess the efficacy of interventions or drug trials, and significantly decrease the manual burden when analysing large cohorts. Clinically, this pipeline can be used to assess longitudinal changes in skeletal muscle mass over time, evaluate the efficacy of pulmonary rehabilitation and assess the effectiveness of therapeutic intervention on skeletal muscle mass (e.g., nutrition, exercise training, anabolic steroids). Future research should develop age, sex and BMI-specific normal reference values and prediction equations for PMA so that researchers and/or clinicians can better assess the (ab)normality of individual measurements. Other potential future directions include automated pectoralis sub-muscle segmentation, and measurements of different muscle groups, such as the paraspinals. Finally, future work should explore the application of this tool in individuals with other respiratory-related diseases where measurements of skeletal muscle have been associated with outcomes, such as ILD [10] and lung cancer [40].
Conclusions
We developed a fully automated pipeline for aortic arch localisation and PMA segmentation from routinely collected chest CT scans. Our pipeline was accurate and highly generalisable to an external never-before-seen set of CT images from a group of clinical research study participants with COPD. Our findings motivate the use of this fully automated pipeline for pectoralis segmentation in both clinical and research settings.
Supplementary material
Supplementary Material
Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author.
Supplementary material 00485-2023.SUPPLEMENT
Acknowledgements
The authors would like to thank Benjamin Smith and Adamo Donovan (McGill University) for their technical input for the PMA segmentations.
Footnotes
Provenance: Submitted article, peer reviewed.
Conflicts of interest: D. Genkin reports no conflicts of interest.
Conflicts of interest: A.R. Jenkins acknowledges salary support from Le Fonds de Recherche du Québec Santé - Postdoctoral Training fellowship.
Conflicts of interest: N. van Noord reports no conflicts of interest.
Conflicts of interest: K. Makimoto reports no conflicts of interest.
Conflicts of interest: S. Collins reports no conflicts of interest.
Conflicts of interest: M.K. Stickland acknowledges salary support from the Canadian Institutes of Health Research, Natural Sciences and Engineering Council of Canada, and University Hospital Foundation.
Conflicts of interest: W.C. Tan reports no conflicts of interest.
Conflicts of interest: J. Bourbeau acknowledges salary support from the Canadian Institute of Health Research, Réseau en santé respiratoire du FRQS, McGill University Health Centre Foundation, AstraZeneca Canada Ltd, Boehringer Ingelheim Canada Ltd, GlaxoSmithKline Canada Ltd, Grifols, Novartis, Sanofi, Trudell Canada Ltd, Pfizer Canada Ltd and COVIS Pharma Canada Ltd.
Conflicts of interest: D. Jensen reports no conflicts of interest.
Conflicts of interest: M. Kirby is a paid consultant at VIDA Diagnostics.
Support statement: D. Genkin acknowledges salary support from the Natural Sciences and Engineering Research Council (NSERC). M. Kirby acknowledges support from the NSERC Discovery Grant, the Early Researchers Award Program and the Canada Research Chair Program (Tier II). A.R. Jenkins acknowledges salary support from Le Fonds de Recherche du Québec Santé - Postdoctoral Training fellowship. D. Jensen holds a Canada Research Chair in Clinical Exercise and Respiratory Physiology (Tier 2) from the Canadian Institutes of Health Research. This study was supported by a Team Project Grant from the Québec Respiratory Health Research Network to D. Jensen as Principal Investigator, and A. Jenkins, M. Kirby and J. Bourbeau as co-investigators. Funding information for this article has been deposited with the Crossref Funder Registry.
Ethics statement: Institutional review board approval was obtained at each site.
- Received July 13, 2023.
- Accepted November 9, 2023.
- Copyright ©The authors 2024
This version is distributed under the terms of the Creative Commons Attribution Non-Commercial Licence 4.0. For commercial reproduction rights and permissions contact permissions{at}ersnet.org
References
