Cardiovascular disease and all-cause mortality risk prediction from abdominal CT using deep learning

Full text

PROCEEDINGS OF SPIE
SPIEDigitalLibrary.org/conference-proceedings-of-spie
Cardiovascular disease and all-cause
mortality risk prediction from
abdominal CT using deep learning
Elton, Daniel, Chen, Andy, Pickhardt, Perry, Summers,
Ronald
Daniel C. Elton, Andy Chen, Perry J. Pickhardt, Ronald M. Summers,
"Cardiovascular disease and all-cause mortality risk prediction from
abdominal CT using deep learning," Proc. SPIE 12033, Medical Imaging
2022: Computer-Aided Diagnosis, 120332N (4 April 2022); doi:
10.1117/12.2612620
Event: SPIE Medical Imaging, 2022, San Diego, California, United States
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 07 Apr 2022 Terms of Use: https://www.spiedigitallibrary.org/terms-of-use

Cardiovascular disease and all-cause mortality risk prediction
from abdominal CT using deep learning
Daniel C. Eltona, Andy Chena, Perry J. Pickhardt∗b, and Ronald M. Summers∗a
aImaging Biomarkers and Computer-Aided Diagnosis Laboratory, Radiology and Imaging
Sciences, National Institutes of Health Clinical Center, Bethesda, MD 20892-1182, USA
bSchool of Medicine and Public Health, University of Wisconsin, Madison, WI 53726, USA
ABSTRACT
Cardiovascular disease is the number one cause of mortality worldwide. Risk prediction can help incentivize
lifestyle changes and inform targeted preventative treatment. In this work we explore utilizing a convolutional
neural network (CNN) to predict cardiovascular disease risk from abdominal CT scans taken for routine CT
colonography in otherwise healthy patients aged 50-65. We find that adding a variational autoencoder (VAE) to
the CNN classifier improves its accuracy for five year survival prediction (AUC 0.787 vs. 0.768). In four-fold cross
validation we obtain an average AUC of 0.787 for predicting five year survival and an AUC of 0.767 for predicting
cardiovascular disease. For five year survival prediction our model is significantly better than the Framingham
Risk Score (AUC 0.688) and of nearly equivalent performance to method demonstrated in Pickhardt et al. (AUC
0.789) which utilized a combination of five CT derived biomarkers.
Keywords: cardiovascular disease, machine learning, deep learning, longevity, artificial intelligence, risk, ab-
dominal CT, all-cause mortality
1. PURPOSE
Globally, cardiovascular disease (CVD) remains the number one cause of mortality, causing 17.9 million deaths
in 2019, including 38% of all premature deaths from noncommunicable diseases.1As standards of living increase
globally, people have more time and resources to invest in preventative measures to protect their long term
health. Since the 1960s, the rate of deaths from CVD in the United States has been reduced by about 50%
through a mixture of lifestyle changes and improved treatment. It has been argued that up to 90% of CVD
is preventable if lifestyle changes and other interventions are implemented early enough.2Informing a patient
that they have higher than average CVD risk can help encourage positive lifestyle changes. Additionally, several
primary prevention treatments are under investigation to reduce CVD risk including metformin,3,4low-dose
aspirin,5and gene therapy.6Given the costs and risks associated with these treatments, targeting them to the
most at-risk patients will be important.
Genetic screening can help with targeting but completely neglects the effects of environmental exposures. A
popular risk scoring system for CVD risk used currently is the Framingham risk score (FRS) which factors in
age, sex, cholesterol level, and blood pressure.7In the Framingham Heart Study 8,491 patients were followed
for 12 years and it was found that the FRS achieved a C-statistic (equivalent to AUC for censored data) for
predicting CVD of 0.763 (95% confidence interval (CI) 0.746 - 0.780) in men and 0.793 (95% CI, 0.772 - 0.814)
in women.8Previously we showed that the FRS yields an AUC of 0.688 (95% CI 0.650–0.727) for predicting five
year survival for the patient cohort we will study.9A 2019 meta-analysis compared three popular risk models
for predicting 10-year CVD risk - the FRS, Framingham Adult Treatment Panel III model, and pooled cohort
model.10 They found modest C-statistics ranging between 0.68-0.74.10
Imaging biomarkers hold much promise for improving CVD risk scoring. In particular, coronary artery
calcification (CAC) scoring has been heavily studied.11,12 CAC scores are typically obtained using a specialized
ECG gated non-contrast cardiac CT scan. The addition of CAC score has been shown to improve risk prediction
∗Co-senior authors. Further author information: send correspondence to Daniel Elton (delton@mgh.harvard.edu) or
Ronald Summers (rms@nih.gov)
Medical Imaging 2022: Computer-Aided Diagnosis, edited by Karen Drukker, Khan M. Iftekharuddin,
Hongbing Lu, Maciej A. Mazurowski, Chisako Muramatsu, Ravi K. Samala, Proc. of SPIE
Vol. 12033, 120332N · © 2022 SPIE · 1605-7422 · doi: 10.1117/12.2612620
Proc. of SPIE Vol. 12033 120332N-1
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 07 Apr 2022
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use

when added to more traditional risk factors such as age, gender, blood biomarkers, and family history.13–15 In the
Multi-Ethnic Study of Atherosclerosis (MESA) dataset the FRS yielded a C-statistic for CVD risk (median follow
up time 7.6 years) of 0.62 but that rose to 0.78 when CAC score was included.13 Work by Mets et al. obtained a
C-statistic of 0.71 (95% CI 0.67-0.76) for three year survival prediction using the National Lung Screening Trial
dataset with a multivariate model that combined CAC score and clinical parameters (age, smoking status, etc.)14
McClelland et al. obtained a C-statistic of 0.80 for 10 year prediction of coronary heart disease when combining
CAC score with blood biomarkers, family history, age, and gender.15 Recently several works have shown how
deep learning systems can automate CAC scoring.16–20 Interestingly, Chao et al. have demonstrated that feeding
a box around the heart from a chest CT into a multi-view convolutional neural net (CNN) architecture can lead
to improved CVD and mortality prediction over the state of the art for automated CAC scoring or manual CAC
grading.20 This suggests that CNNs can extract additional features beyond the coronary artery calcification
which help inform the risk prediction. Partially by inspired that work, in this work we study feeding an entire
abdominal CT scan into a CNN to perform risk scoring rather than extracting custom biomarkers such as aortic
calcification score and visceral/subcutaneous fat ratio like we did in a prior work.9
The vast majority of prior work on CVD or all-cause mortality risk prediction has looked at either chest
CT19–24 or chest X-ray.25,26 Among the works that utilize chest CT, the publicly available National Lung
Screening Trial (NLST) dataset has been heavily utilized14,19,20,23,24,26 in addition to the Dutch-Belgian lung
cancer screening trial,21 or in-house chest CT datasets.22 NLST participants have a history of smoking, putting
them at higher than average risk for CVD. In contrast, our dataset consists of scans from an otherwise healthy
patient cohort. In this work we focus on risk scoring from abdominal CT scans, which has only been studied
previously in a handful of works.9,27,28 Abdominal CT are rich with biomarkers known to be relevant to
cardiovascular risk such as visceral fat and aortic plaque.29,30 This is the first work to feed the entire abdominal
CT scan into a deep learning model.
μ
σ
Sampled latent vector
(size = 256)
VAE architecture #2
Hidden fully connected layer
(size = 512)
Output branch for CNN / VAE architecture #1
Fully connected
feedforward
network
3D Convolution
Classification
sigmoid output
unit
16
32
64
64
64
Resampled
3D CT scan
volume
VAE Decode r (op tiona l)
Figure 1. The architectures employed. The CNN architecture consists of five CNN layers followed by two fully connected
layers. Optionally a VAE decoder may be attached to the CNN to encourage the CNN to obtain good high level features. If
the VAE is attached, one can experiment with attaching the classification output to the vector of latent means (architecture
2)
2. METHODS
The dataset we utilize consists of CT colonography (CTC) scans from the University of Wisconsin Medical
Center.27 9,223 people underwent CTC scans between April 2004 and December 2016. Further details on
Proc. of SPIE Vol. 12033 120332N-2
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 07 Apr 2022
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use

4826 scans 255 scans 1694 scans
6775 Patients with 5 year follow-
up data, one scan
per patient
Training set
(71.23%)
Testing set
(25%)
Validation set
(3.78%)
9393 patients who underwent CT
Colonographies at
University of Wisconsin-Madison
Excluded:
170 scans either corrupted or
not supine scans
2448 patients without 5 year
follow-up data
Figure 2. Patient flow chart showing how the patient cohort was selected and split into train, validation, and test sets
for five year survival prediction.
the cohort are provided in Pickhardt et al, 2020.9As in prior works9,27 cardiovascular disease was defined as
myocardial infarction, cerebrovascular accident, or development of congestive heart failure. These reflect the
endpoints considered by the FRS for cardiovascular disease.
We found 6775 patients had five year follow up data indicating if they survived for five years and out of those
216 (3%) died within five years. Similarly, 7008 patients had follow up data indicating whether or not they were
diagnosed with CVD within five years, and out of those 399 developed CVD (5.6%). We used roughly 71% of
the data for training, 5% for validation, and 25% for test. The data flow and number of patients in the training,
validation, and testing folds for 5 year survival prediction is summarized in figure 2.
The architectures for the CNN and VAE are shown in figure 1. The two architectures were identical apart
from the presence of the VAE decoder. To reduce the number of parameters in the network the first convolutional
layer contains a large 7x7x7 kernel applied with a stride of 2 in each direction and padding of 3 on the edges. The
rest of the convolutional layers use 4x4x4 kernels with a stride of 2 and padding of 1. Group normalization and
dropout (dropout rate 0.5) were used after each convolutional layer. In summary, the CNN encoder contains 5
convolutional layers and two feedforward networks and has a total of 5,414,558 parameters. The VAE architecture
contains 3 additional fully connected layers and 5 additional upsampling layers and has a total of 21,060,930
parameters. We use a relatively small latent vector size of D= 256 to encourage the network to find efficient
high level features.
We also attempt to add the patient age and gender information to the model. As is standard practice in
machine learning, we encode the age and gender information in one-hot vectors. Ages were encoded in a vector
of length 60 corresponding to ages between 30-90 years old (any patient with age less than 30 is set to 30 and
any above 90 is set to 90). The one hot vectors are concatenated to the hidden layer.
The images are resampled to a size of 192x192x128 using trilinear interpolation, clipped to [-500, 500], and
then rescaled to between [0, 1]. To deal with the data imbalance (only ≈3% of patients died with 5 years) we
reweight the data during training so the ratio of survived/died is 50/50 during training. The RAdam optimizer31
was used with a learning rate of 0.001 and batch size of 4. The data augmentations employed were limited to
random rotations around a random axis by +/- 0-15 degrees, random 0-20% cropping in the cranial-caudal
direction, and random left-right flipping.
Most existing works using a VAE, such as van Velzen et al.,23 train an autoencoder first for image recon-
struction and then train a separate classification model that utilizes either the latent vector or a hidden layer
vector taken from the “bottleneck” portion of the VAE. In this work we perform joint training, training our
Proc. of SPIE Vol. 12033 120332N-3
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 07 Apr 2022
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use

VAE model for both image reconstruction and classification at the same time. This procedure is inspired by
previous work where a VAE decoder was attached to a 3D V-Net.32 The author claims that adding the VAE
helps regularize the 3D V-Net and prevent overfitting.32 In a similar way we hypothesize that joint training with
the VAE attached will improve the generalization performance of the classifier by encouraging the network to
learn robust high level representations. Joint training is more challenging to implement since it requires carefully
weighting the loss terms, otherwise one or more loss terms may be ignored during gradient descent training.
The loss function applied at iteration ifor the VAE architecture was:
Li=βLCE +LL1 + ΓiγLKL (1)
where
LL1 =||Iinput −Ipred|| LKL =1
NXµ2+σ2−log σ2−1 (2)
The loss function for the classifier was the standard cross-entropy loss weighted by a factor β. For the VAE
reconstruction loss we used an L1 loss function. As has been adopted in prior works,33 a weighting factor Γiwas
applied to Kullback-Leibler loss term of the VAE and linearly increased from 0 to 1 over the course of the first
5,000 iterations. This practice prevents the optimizer from only focusing on minimizing the KL loss in the early
stages of training. The factors βand γhad to be determined empirically. After experimenting with weightings
of β∈ {1,10,100}and γ∈ {0.001,0.0001}we used β= 10 and γ= 0.001 as this was the only combination where
all three loss functions decreased during training. Further tuning of these hyperparameters was not attempted.
The F1 score and AUC in the validation set were monitored during training. All models were trained for exactly
four epochs, which was found to be sufficient for the validation F1 and AUC to converge.
3. RESULTS
AUC 5-year AUC 5-year AUC 5-year
method mortality CVD or mortality CVD
CNN only 0.768±0.026
CNN+VAE (arch #1) 0.787±0.030 0.767 ±0.036
CNN+VAE (arch #1) + age+sex 0.777±0.031
CNN+VAE (arch #2) 0.770±0.034
FRS90.688 0.695
BMI90.499 0.552
best CT derived model from Pickhardt et al. 202090.789 0.742
best multivariate model (CT biomarkers+FRS)90.796 0.751
Table 1. Summary of average AUCs in four-fold cross validation and standard deviation over the four folds.
box size validation AUC test AUC
128x128x128 0.75 0.74
192x192x128 0.76 0.76
192x192x192 0.78 0.79
256x256x128 0.78 0.76
Table 2. Hyperparameter optimization experiments testing different box sizes for five year survival prediction using the
CNN-only architecture. Validation AUC is an average of the validation AUCs obtained in the the last epoch of training
and the test AUC is the AUC on the hold-out test set for the final model after 4 epochs of training.
A summary of the average AUCs obtained in four-fold cross validation is provided in table 3and select ROC
curves are shown in figure 4. In line with our hypothesis, we found that the VAE obtained a higher AUC than the
CNN only (0.787 vs 0.768), although we note that due to the limited number of folds employed means the result is
not statistically significant (Welch’s t-test p= 0.34). VAE architecture option #2, where the classifier is attached
to the latent vector means, performed slightly worse than VAE architecture option #1. The concatenation of
age and sex information to the final layer of the classifier did not improve the AUC. This may partially be due
Proc. of SPIE Vol. 12033 120332N-4
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 07 Apr 2022
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use

Figure 3. Visualization of a case from the publicly available Kidney Tumor Segmentation Challenge (KiTS) dataset which
contains aortic plaque. The patterns observed here were found to be remarkably consistent across five different cases we
looked at. This method is likely showing only where the network is looking generally in the first few layers and it is hard
to draw many firm conclusions about how the model functions internally on the basis of this sort of visualization.
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
1 - specificity (false positive rate)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
sensitivity
CNN only, avg AUC = 0.77
CNN+VAE, avg AUC = 0.79
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
1 - specificity (false positive rate)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
sensitivity
CNN+VAE CVD only, avg AUC = 0.77
Figure 4. Select average ROC curves under 4-fold cross validation for five year survival prediction (left) and five year
CVD prediction (right). The shaded regions show the standard deviation over the four folds.
to the fact that the ages of the patients did not vary that much. It may be possible to better incorporate the
age and gender information, eg. by adding an additional fully connected layer.
The results in table 3were obtained using a box size of 192x192x128. After obtaining those results we ran
some hyperparameter optimization experiments to investigate the effect of box size, the results of which are
shown in table 3). The results show a slightly lower AUC is found when using a smaller box of size 128x128x128
and a slightly higher AUC is obtained when the box size is increased to 192x192x192, at least for the one fold
under consideration. The results suggest that a larger box of 192x192x192 may be beneficial but there is little
benefit to increasing the box size beyond that.
We experimented with visualizing model using guided backpropagation34 using a publicly available code-
base.35 An example of the visualization obtained can be found in figure 3. As shown by Adebayo et al.,36
“saliency” methods like guided backprop are only sensitive to what is going on in the first few layers. Such
methods show roughly where the network is looking, but do not provide much insight into what the network is
doing and should be interpreted with caution.37,38
Proc. of SPIE Vol. 12033 120332N-5
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 07 Apr 2022
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use

4. NEW OR BREAKTHROUGH WORK
For the first time we have shown how a deep learning model can predict all-cause mortality risk directly from
an abdominal CT using a novel jointly trained VAE architecture. We found that the use of the VAE decoder
improves the model over using just a plain CNN architecture. The VAE based model is significantly better
than using the Framingham Risk Score at predicting five year mortality and is of equivalent accuracy to the
best system for abdominal CT demonstrated in Pickhardt et al.9(AUC 0.787 vs 0.789). Our model has the
advantage of being much simpler than the prior system, which requires the use of five independent software tools
for plaque, bone mineral density, visceral/subcutaneous fat, liver fat, and muscle quantification. On a modern
workstation with a NVIDIA Quadro RTX 8000 GPU our model runs in about four seconds and that time reduces
to less than two seconds when our model is preloaded into GPU memory. Many avenues are open for improving
our model further.
ACKNOWLEDGMENTS
This research was supported in part by the Intramural Research Program of the National Institutes of Health
Clinical Center.
REFERENCES
[1] “Cardiovascular diseases (CVDs) https://www.who.int/news-room/fact-sheets/detail/
cardiovascular-diseases-(cvds),” (June 2021).
[2] McGill, H. C., McMahan, C. A., and Gidding, S. S., “Preventing heart disease in the 21st century,” Circu-
lation 117, 1216–1227 (Mar. 2008).
[3] Han, Y., Xie, H., Liu, Y., Gao, P., Yang, X., and Shen, Z., “Effect of metformin on all-cause and cardiovascu-
lar mortality in patients with coronary artery diseases: a systematic review and an updated meta-analysis,”
Cardiovascular Diabetology 18 (July 2019).
[4] Kulkarni, A. S., Gubbi, S., and Barzilai, N., “Benefits of metformin in attenuating the hallmarks of aging,”
Cell Metabolism 32, 15–30 (July 2020).
[5] Patrono, C. and Baigent, C., “Role of aspirin in primary prevention of cardiovascular disease,” Nature
Reviews Cardiology 16, 675–686 (June 2019).
[6] King, A., “A CRISPR edit for heart disease,” Nature 555, S23–S25 (Mar. 2018).
[7] Wilson, P. W. F., D’Agostino, R. B., Levy, D., Belanger, A. M., Silbershatz, H., and Kannel, W. B.,
“Prediction of coronary heart disease using risk factor categories,” Circulation 97, 1837–1847 (May 1998).
[8] D’Agostino, R. B., Vasan, R. S., Pencina, M. J., Wolf, P. A., Cobain, M., Massaro, J. M., and Kannel,
W. B., “General cardiovascular risk profile for use in primary care,” Circulation 117, 743–753 (Feb. 2008).
[9] Pickhardt, P. J., Graffy, P. M., Zea, R., Lee, S. J., Liu, J., Sandfort, V., and Summers, R. M., “Automated
CT biomarkers for opportunistic prediction of future cardiovascular events and mortality in an asymptomatic
screening population: a retrospective cohort study,” The Lancet Digital Health 2, e192–e200 (Apr. 2020).
[10] Damen, J. A., Pajouheshnia, R., Heus, P., Moons, K. G. M., Reitsma, J. B., Scholten, R. J. P. M., Hooft,
L., and Debray, T. P. A., “Performance of the framingham risk models and pooled cohort equations for
predicting 10-year risk of cardiovascular disease: a systematic review and meta-analysis,” BMC Medicine 17
(June 2019).
[11] McClelland, R. L., Nasir, K., Budoff, M., Blumenthal, R. S., and Kronmal, R. A., “Arterial age as a function
of coronary artery calcium (from the multi-ethnic study of atherosclerosis [MESA]),” The American Journal
of Cardiology 103, 59–63 (Jan. 2009).
[12] Chiles, C., Duan, F., Gladish, G. W., Ravenel, J. G., Baginski, S. G., Snyder, B. S., DeMello, S., Desjardins,
S. S., and and, R. F. M., “Association of coronary artery calcification and mortality in the national lung
screening trial: A comparison of three scoring methods,” Radiology 276, 82–90 (July 2015).
[13] Yeboah, J., McClelland, R. L., Polonsky, T. S., Burke, G. L., Sibley, C. T., O’Leary, D., Carr, J. J.,
Goff, D. C., Greenland, P., and Herrington, D. M., “Comparison of novel risk markers for improvement in
cardiovascular risk assessment in intermediate-risk individuals,” JAMA 308, 788 (Aug. 2012).
Proc. of SPIE Vol. 12033 120332N-6
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 07 Apr 2022
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use

[14] Mets, O. M., Vliegenthart, R., Gondrie, M. J., Viergever, M. A., Oudkerk, M., de Koning, H. J., Mali, W. P.,
Prokop, M., van Klaveren, R. J., van der Graaf, Y., Buckens, C. F., Zanen, P., Lammers, J.-W. J., Groen,
H. J., Isgum, I., and de Jong, P. A., “Lung cancer screening CT-based prediction of cardiovascular events,”
JACC: Cardiovascular Imaging 6, 899–907 (Aug. 2013).
[15] McClelland, R. L., Jorgensen, N. W., Budoff, M., Blaha, M. J., Post, W. S., Kronmal, R. A., Bild, D. E.,
Shea, S., Liu, K., Watson, K. E., Folsom, A. R., Khera, A., Ayers, C., Mahabadi, A.-A., Lehmann, N.,
J¨ockel, K.-H., Moebus, S., Carr, J. J., Erbel, R., and Burke, G. L., “10-year coronary heart disease risk pre-
diction using coronary artery calcium and traditional risk factors,” Journal of the American College of
Cardiology 66, 1643–1653 (Oct. 2015).
[16] Gonz´alez, G., Washko, G. R., Est´epar, R. S. J., Cazorla, M., and Espinosa, C. C., “Automated agatston
score computation in non-ECG gated CT scans using deep learning,” in [Medical Imaging 2018: Image
Processing], Angelini, E. D. and Landman, B. A., eds., SPIE (Mar. 2018).
[17] Commandeur, F., Slomka, P. J., Goeller, M., Chen, X., Cadet, S., Razipour, A., McElhinney, P., Gransar,
H., Cantu, S., Miller, R. J. H., Rozanski, A., Achenbach, S., Tamarappoo, B. K., Berman, D. S., and Dey, D.,
“Machine learning to predict the long-term risk of myocardial infarction and cardiac death based on clinical
risk, coronary calcium, and epicardial adipose tissue: a prospective study,” Cardiovascular Research 116,
2216–2225 (Dec. 2019).
[18] Lee, H., Martin, S., Burt, J. R., Bagherzadeh, P. S., Rapaka, S., Gray, H. N., Leonard, T. J., Schwemmer, C.,
and Schoepf, U. J., “Machine learning and coronary artery calcium scoring,” Current Cardiology Reports 22
(July 2020).
[19] Zeleznik, R., Foldyna, B., Eslami, P., Weiss, J., Alexander, I., Taron, J., Parmar, C., Alvi, R. M., Banerji,
D., Uno, M., Kikuchi, Y., Karady, J., Zhang, L., Scholtz, J.-E., Mayrhofer, T., Lyass, A., Mahoney, T. F.,
Massaro, J. M., Vasan, R. S., Douglas, P. S., Hoffmann, U., Lu, M. T., and Aerts, H. J. W. L., “Deep
convolutional neural networks to predict cardiovascular risk from computed tomography,” Nature Commu-
nications 12 (Jan. 2021).
[20] Chao, H., Shan, H., Homayounieh, F., Singh, R., Khera, R. D., Guo, H., Su, T., Wang, G., Kalra, M. K., and
Yan, P., “Deep learning predicts cardiovascular disease risks from lung cancer screening low dose computed
tomography,” Nature Communications 12 (May 2021).
[21] de Vos, B. D., de Jong, P. A., Wolterink, J. M., Vliegenthart, R., Wielingen, G. V., Viergever, M. A., and
Iˇsgum, I., “Automatic machine learning based prediction of cardiovascular events in lung cancer screening
data,” in [Medical Imaging 2015: Computer-Aided Diagnosis], Hadjiiski, L. M. and Tourassi, G. D., eds.,
SPIE (Mar. 2015).
[22] Oakden-Rayner, L., Carneiro, G., Bessen, T., Nascimento, J. C., Bradley, A. P., and Palmer, L. J., “Pre-
cision radiology: Predicting longevity using feature engineering and deep learning methods in a radiomics
framework,” Scientific Reports 7(May 2017).
[23] van Velzen, S., Zreik, M., Lessmann, N., Viergever, M. A., de Jong, P. A., Verkooijen, H. M., and Iˇsgum,
I., “Direct prediction of cardiovascular mortality from low-dose chest CT using deep learning,” in [Medical
Imaging 2019: Image Processing], Angelini, E. D. and Landman, B. A., eds., SPIE (Mar. 2019).
[24] Guo, H., Kruger, U., Wang, G., Kalra, M. K., and Yan, P., “Knowledge-based analysis for mortality
prediction from CT images,” IEEE Journal of Biomedical and Health Informatics 24, 457–464 (Feb. 2020).
[25] Karargyris, A., Kashyap, S., Wu, J. T., Sharma, A., Moradi, M., and Syeda-Mahmood, T., “Age prediction
using a large chest x-ray dataset,” in [Medical Imaging 2019: Computer-Aided Diagnosis], Hahn, H. K. and
Mori, K., eds., SPIE (Mar. 2019).
[26] Raghu, V. K., Weiss, J., Hoffmann, U., Aerts, H. J., and Lu, M. T., “Deep learning to estimate biological
age from chest radiographs,” JACC: Cardiovascular Imaging (Mar. 2021).
[27] O’Connor, S. D., Graffy, P. M., Zea, R., and Pickhardt, P. J., “Does nonenhanced CT-based quantification
of abdominal aortic calcification outperform the framingham risk score in predicting cardiovascular events
in asymptomatic adults?,” Radiology 290, 108–115 (Jan. 2019).
[28] Zambrano Chaves, J. M., Chaudhari, A. S., Wentland, A. L., Desai, A. D., Banerjee, I., Boutin, R. D.,
Maron, D. J., Rodriguez, F., Sandhu, A. T., Jeffrey, R. B., Rubin, D., and Patel, B., “Opportunistic
assessment of ischemic heart disease risk using abdominopelvic computed tomography and medical record
data: a multimodal explainable artificial intelligence approach,” medRxiv (2021).
Proc. of SPIE Vol. 12033 120332N-7
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 07 Apr 2022
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use

[29] Sethi, A., Taylor, L., Ruby, J. G., Venkataraman, J., Sorokin, E., Cule, M., and Melamud, E., “Calcification
of abdominal aorta is a high risk underappreciated cardiovascular disease factor in a general population,”
medRxiv (2020).
[30] Pickhardt, P. J., Graffy, P. M., Perez, A. A., Lubner, M. G., Elton, D. C., and Summers, R. M., “Opportunis-
tic screening at abdominal CT: Use of automated body composition biomarkers for added cardiometabolic
value,” RadioGraphics 41, 524–542 (Mar. 2021).
[31] Liu, L., Jiang, H., He, P., Chen, W., Liu, X., Gao, J., and Han, J., “On the variance of the adaptive
learning rate and beyond,” in [Proceedings of the 8th International Conference on Learning Representations
(ICLR)], (2020).
[32] Myronenko, A., “3D MRI brain tumor segmentation using autoencoder regularization,” arXiv e-prints ,
arXiv:1810.11654 (Oct. 2018).
[33] Sandfort, V., Yan, K., Graffy, P. M., Pickhardt, P. J., and Summers, R. M., “Use of variational autoen-
coders with unsupervised learning to detect incorrect organ segmentations at CT,” Radiology: Artificial
Intelligence 3, e200218 (July 2021).
[34] Springenberg, J. T., Dosovitskiy, A., Brox, T., and Riedmiller, M. A., “Striving for simplicity: The all
convolutional net,” in [3rd International Conference on Learning Representations, ICLR 2015, San Diego,
CA, USA, May 7-9, 2015, Workshop Track Proceedings ], Bengio, Y. and LeCun, Y., eds. (2015).
[35] B¨ohle, M., Eitel, F., Weygandt, M., and Ritter, K., “Layer-wise relevance propagation for explaining
deep neural network decisions in MRI-based alzheimer's disease classification,” Frontiers in Aging Neu-
roscience 11 (July 2019).
[36] Adebayo, J., Gilmer, J., Muelly, M., Goodfellow, I., Hardt, M., and Kim, B., “Sanity checks for saliency
maps,” in [Proceedings of the 32nd International Conference on Neural Information Processing Systems],
NIPS’18, 9525–9536, Curran Associates Inc., Red Hook, NY, USA (2018).
[37] Rudin, C., “Stop explaining black box machine learning models for high stakes decisions and use inter-
pretable models instead,” Nature Machine Intelligence 1, 206–215 (May 2019).
[38] Ghassemi, M., Oakden-Rayner, L., and Beam, A. L., “The false hope of current approaches to explainable
artificial intelligence in health care,” The Lancet Digital Health 3, e745–e750 (Nov. 2021).
Proc. of SPIE Vol. 12033 120332N-8
Downloaded From: https://www.spiedigitallibrary.org/conference-proceedings-of-spie on 07 Apr 2022
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use