Enhanced Multimodal Timely Prediction of Pulmonary Fibrosis

Progression with Uncertainty Estimation from Chest CT Images and

Clinical Metadata

Mohamed Dahmane

Computer Research Institute of Montreal (CRIM),

405 Av. Ogilvy #101, Montreal, Quebec, Canada

Keywords:

CT Scans, Uncertainty Estimation, Pulmonary Fibrosis, Multimodal Deep Learning, Clinical Metadata.

Abstract:

Pulmonary Fibrosis (PF) is a progressive chronic illness in which the lung tissues become increasingly scarred

and damaged, leading to irreversible loss of their capacity to oxygenate vital organs. The speciﬁc causes of

the illness are often unknown in many cases. Assessment of the severity of the lung disease is critical for

physicians to diagnose PF early, control disease decline, and manage damage progression. The Forced Vital

Capacity (FVC) of the lungs measured by a spirometer, is a good indicator of the severity of the condition of

the lungs. In this work, we investigated an approach for early diagnosis of PF and showcased a multimodal

architecture that predicts the FVC of patients at different stages of the disease. We propose an anti-Elu interme-

diate block and an anti-Relu conﬁdence block to predict the pulmonary ﬁbrosis progression. The uncertainty

estimation block proved effective in predicting the FVC using data from initial spirometry measurements, clin-

ical meta-data and CT images. Evaluation of the model on the OSIC pulmonary ﬁbrosis progression dataset

showed improved performance compared to state-of-the-art methods, with an average modiﬁed Laplace log-

likelihood score of -6.8227 on a private test set.

1 INTRODUCTION

Pulmonary Fibrosis (PF) is often characterized by dif-

ﬁculty breathing, which can progressively worsen un-

til the lungs are no longer able to supply vital organs

with sufﬁcient oxygen. In some cases, this damage

can lead to serious health conditions such as Progres-

sive Fibrosis Interstitial Lung Disease (PID) or Idio-

pathic Pulmonary Fibrosis (IPF). Timely and accurate

diagnosis of the stage of pulmonary ﬁbrosis is essen-

tial in reducing the burden of morbidity and mortality

related to lung diseases. Chest imaging, such as X-ray

and high-resolution computed tomography (HRCT),

is one means of diagnosing PF, as well as other tests

and procedures used by radiologists. However, accu-

rately diagnosing PF, particularly predicting the stage

of a progressive disease like PID, can be challenging.

Radiological imaging provides a dedicated tool for vi-

sually assessing the presence of ﬁbrotic tissue and de-

termining the development of lung scarring.

Several research works propose imaging diagno-

sis approaches to aid radiologists in diagnosing lung

https://orcid.org/0000-0002-2670-1433

diseases. Various consortiums worldwide, such as the

Open Source Imaging Consortium (OSIC), Radiolog-

ical Society of North America (RSNA), and Society

of Thoracic Radiology (STR), bring together clinical

researchers and data scientists to improve radiology-

based imaging through deep learning and artiﬁcial

intelligence. These organizations collect extensive

datasets of high-resolution CT images and relevant

metadata to develop advanced multimodal solutions.

In the literature, many research works have pro-

posed using CT scans as a unimodal source of in-

formation to assess the evolution of Idiopathic Pul-

monary Fibrosis. However, few studies have explored

predicting the disease progression from multimodal

data by predicting the Forced Vital Capacity (FVC),

which is an important indicator of pulmonary function

in IPF (du Bois et al., 2011). Moreover, in this work

we enhanced the the uncertainty estimation by intro-

ducing a new anti-Relu block. The paper is organized

as follows: Section 2 discusses the related works in

computer-assisted ﬁbrotic lung disease assessment. In

section 3, we investigate the OSIC pulmonary ﬁbro-

sis data used to evaluate our approach. Section 4 pro-

vides a detailed description of the methodology. Sec-

Dahmane, M.

Enhanced Multimodal Timely Prediction of Pulmonary Fibrosis Progression with Uncertainty Estimation from Chest CT Images and Clinical Metadata.

DOI: 10.5220/0012304100003660

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 19th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2024) - Volume 4: VISAPP, pages

461-468

ISBN: 978-989-758-679-8; ISSN: 2184-4321

461

tion 5, presents the evaluation and the performance of

the proposed approach. Finally, section 6 gives con-

clusions and some insights on potential direction and

future works.

2 RELATED WORK

In recent years, there has been an increasing trend

in the prevalence of idiopathic pulmonary ﬁbrosis

(IPF) in the USA (Nalysnyk et al., 2012). In a study

by (Raghu et al., 2014) on US Medicare patients aged

65 years and older, the authors found that patients

with a median age of 79.4±7.2 years had a survival

time of 3.8 years. The study also revealed an overall

incidence ratio of 93.7 cases per 100k persons. An-

other study by (Hutchinson et al., 2015) reported an

increase in mortality due to pulmonary ﬁbrosis, rang-

ing from 4.64 per 100k for Spain to 8.28 for England

and Wales.

The integration of assitive vision and AI in health-

care has become essential due to the abundance of

clinical data. CT scans are particularly useful for

visually estimating the stage of lung deterioration.

A clinical association was found between interstitial

lung abnormalities (ILA) judged by high-resolution

computed tomography (HRCT) and idiopathic pul-

monary ﬁbrosis (IPF), as reported in (Wells and

Kokosi, 2016; Scatarige et al., 2003).

Motivated by recent advances in deep learning and

computer vision, researchers have investigated vari-

ous architectures for interstitial lung disease (IDL) de-

tection, segmentation, and classiﬁcation (Soffer et al.,

2022). For instance, in a study by Walsh et al. (Walsh

et al., 2018), deep learning was used to classify ﬁ-

brotic lung diseases from high-resolution CT scans,

and the algorithm outperformed 60 out of 91 radiol-

ogists with a median accuracy of 73.3% compared to

the physicians’ accuracy of 70%. In another study,

Comelli et al. (Comelli et al., 2020) evaluated the

UNet and E-Net segmentation models on 10 patients

with idiopathic pulmonary ﬁbrosis (IPF) and achieved

a segmentation accuracy of 96% using the dice simi-

larity coefﬁcient without any radiologist intervention.

Kido et al. (Kido et al., 2022) developed a deep

neural-network architecture for three-dimensional

segmentation of lung nodules for lung cancer diag-

nosis from CT images. The 3D UNet model’s per-

formance was comparable to human experts with a

dice similarity coefﬁcient of 84.5% and 82.2%, re-

spectively. The authors found that traditional machine

learning techniques such as watershed and graph-cut

provided lower accuracy compared to neural-network

based models, with only 62.8% and 56.6% dice sim-

ilarity coefﬁcient, respectively. In another study

by (Christe et al., 2019), an integrated computer-aided

diagnosis system for IPF was developed using deep

learning on CT images. The system’s performance

was similar to that of radiologists under certain evalu-

ation criteria. The study conducted by (Zucker et al.,

2020) utilized a DCNN model based on ResNet-18

to predict Brasﬁeld scores, which are indicative of

various lung function features such as air trapping,

linear markings, nodular cystic lesions, large lesions,

and overall severity. The authors reported minimal

differences between the model’s Brasﬁeld scores and

those of the experts, except for the large lesion fea-

tures, which had an average Spearman correlation

of only 32% between the model and the radiolo-

gists. However, the correlation rate for large lesion

scores showed a higher rate of 80.2%. In (Agarwala

et al., 2020), a convolutional neural network was ﬁrst

trained on natural images and then ﬁne-tuned on CT

images to automatically segment interstitial lung dis-

ease (ILD) patterns such as emphysema, consolida-

tion, and ﬁbrosis. The reported results were accept-

able, with a classiﬁcation rate of 90% for ﬁbrosis pat-

tern segmentation.

Several research studies based their works on the

OSIC data (Osic, 2023) to predict the decline in

lung function severity which is assessed by measur-

ing the forced vital capacity using a spirometer (Wat-

ters et al., 1986; Noth et al., 2021). The best perfor-

mance was obtained using a bimodal deep learning

model to process CT images and a neural net regres-

sor to process patient clinical metadata. The objec-

tive function was optimized using a multiple quantile

loss function. Efﬁcient-Net was adopted as a back-

bone to process the images. In their study, (Wong

et al., 2021) developed Fibros-Net, an architecture

designed to predict ﬁbrosis progression from chest

scans. The model used CT images, spirometry mea-

surements, and patient clinical metadata to estimate

forced vital capacity (FVC) over a speciﬁc time in-

terval from the OSIC data. The model achieved a

good Laplace log-likelihood score of -6.8188. In con-

trast, FVC-Net from (Yadav et al., 2022) represents

a different architecture that estimates FVC from de-

rived honeycombing features, CT scans, and meta-

data of the OSIC dataset. The model showed a higher

Laplace log-likelihood coefﬁcient of -6.64. A study

by (Mandal et al., 2020) compared the performance

of machine learning models with CNN architecture

in predicting FVC from CT images and patient meta-

data. The experiments showcased good results using

an Elastic-Net regression method achieving a higher

likelihood score of -6.73 on the OSIC dataset.

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

462

3 OSIC PULMONARY FIBROSIS

PROGRESSION DATA

The data used in this study comprises CSV metadata

and CT scans from the OSIC-Pulmonary Fibrosis Pro-

gression dataset (Osic, 2023). The CSV ﬁle contains

metadata information such as age, sex, weeks, smok-

ing status, and FVC percent, which is deﬁned as the

percentage of the typical FVC measure for a person

with similar characteristics. Each patient has a unique

ID, and a baseline chest CT scan is available corre-

sponding to the reference visiting week, i.e., week=0.

The FVC of the patient is measured as a function of

the week number during the follow-up visits, which

span approximately 1 to 2 years. Pre/post visits re-

lating to the reference week are referred to by nega-

tive/positive values. Table 1 provides an example of

the metadata for a particular patient.

Table 1: FVC ground-truth per week and metadata of pa-

tient ID00007637202177411956430.

The task at hand is to predict the Forced Vi-

tal Capacity of a group of patients for expected

weeks. There are 176 unique patients with an aver-

age of 9 visits per person, which occur at different

weeks throughout the year. Figure 1 presents a his-

togram of the number of visiting weeks referred to as

the number of instances per patient. Figure 2, depicts

the distribution of ‘smoking status’ vs. ‘gender’ of the

patients in the dataset. Male ex-smokers are the most

prevalent group, while there are more never-smoked

cases among female patients. The proportion of non-

smokers is almost equal for both sexes. The ﬁgure

highlights the importance of smoking as a risk factor

for pulmonary diseases such as ﬁbrosis.

Figure 1: Instance distribution of unique patients.

The majority of the patients are around 67 years

old with a roughly equal distribution of males and fe-

Figure 2: Instance distribution over smoking status and gen-

der.

males as shown in Figure 3. This is consistent with

the fact that, like many other diseases, age is a major

risk factor for the incidence and prevalence of ﬁbro-

sis. Patients over 60 years old are more likely to be

diagnosed with ﬁbrosis (Raghu et al., 2006).

Figure 3: Distribution of the age of the patients by their

gender.

In ﬁgure 4, the FVC distribution across ‘Smoking

status’ and ‘Gender’ is shown. The measurements are

spread over different intervals, and the average FVC

value of male patients is higher than the correspond-

ing female average for each smoking status, which is

in line with expectations. Figure 5 depicts 28 slices

of the chest Computerized Tomography scan of a par-

ticular patient. Notice that he number of images may

differ for each person.

In the next section, the methodology is covered,

and the predictive models of pulmonary ﬁbrosis pro-

gression are described.

4 MULTIPLE QUANTILE

REGRESSION-BASED MODELS

To estimate the FVC value, we propose a multi-

modal model that incorporates computed tomogra-

phy scans, demographic, and clinical data, including

‘Age’, ‘FVC percent’, ‘Sex’, ‘Week number’, and

‘Smoking status’. First, model

a clinical data-

based model using only clinical information. Next,

Enhanced Multimodal Timely Prediction of Pulmonary Fibrosis Progression with Uncertainty Estimation from Chest CT Images and

Clinical Metadata

463

Figure 4: The measured Forced Vital Capacity distribution across smoking-status and gender.

Figure 5: Chest CT-scan from patient

ID00007637202177411956430.

the multimodal model

CL-CT

is trained and evaluated

on both clinical and CT images. Finally, model

Blend

a blended model predicts the ﬁnal FVC value and the

conﬁdence interval for each week of the considered

period. The performance of the models is analyzed

using the Laplace Log Likelihood (LLL) evaluation

metric, which considers the uncertainty when evalu-

ating the accuracy of the predictions. The problem is

approached as multiple quantile regression problem.

4.1 Clinical Data-Based Model

The clinical data-based model model

depends only

on the clinical data. We applied our conﬁdence block

to the baseline model from (Lhagiimn, 2023). The

input variables of the model consist of both contin-

uous and categorical features. The continuous fea-

tures includes ‘Male’, ‘Female’, ‘Ex-smoker’, ‘Never

smoked’, ‘Currently smokes’, ‘Age’, ‘Week’, ‘Base’,

‘Min percent norm’, and ‘Age week norm’, while the

categorical features consist of ‘Gender’ and ‘Smok-

ing status’. The model incorporates four sequential

blocks and two input blocks. The categorical input

version of the variables ‘Gender’ and ‘Smoking sta-

tus’ are fed to the embedding layer of the prepro-

cessing block for encoding whereas the continuous

variables are fed directly to the self-attention block

with the encoded variables. The conﬁdence block

is plugged to the bottom component of the model

(Fig. 9). As a quantile regressor, the model estimates

the quantiles of a dependent variable expressed as the

conditional median, low quantile, and high quantile

respectively at 50%, 20%, and 80%. Hence, the esti-

mation of the conﬁdence interval is given in general as

a concatenation of the outputs of 3 linear layers since

the model (Fig. 6) has to predict a conﬁdence range

[FVC

, FVC

] as in equation 1.

Figure 6: A conventional conﬁdence block.

FVC

= Linear(o

block B

)

FVC

= Linear(o

block B

) (1)

FVC

= Linear(o

block B

)

We applied a anti-Elu layer (Eq. 2) to the output of

intermediate block A (see Fig. 9). That handles both

negative and positive activation with doubled output

dimension (Fig. 7) compared to the traditional Elu

layer (Eq. 3).

−

= −Elu(−Linear(o

block A

))

= −Elu(Linear(o

block A

)) (2)

o = Concat(o

−

, o

)

o = Elu(Linear(o

block A

)) (3)

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

464

Figure 7: The anti-Elu block.

In comparison to a conventional conﬁdence block, our

conﬁdence block (Fig. 8) uses anti-Relu layers to give

the ﬁnal estimate of the FVC in block C (see Fig. 9).

As deﬁned in (Eq. 4), the layer helps to enforce the

inequality constraint of equation 5.

Figure 8: The anti-Relu-based conﬁdence block.

x = Linear( Concat(o

, o

))

y = Linear(Concat(o

, o

))

δ = Linear(Concat(o

, o

))

−

= −Relu(−δ) (4)

= Relu(δ)

FVC = Concat(x + δ

−

, y, y + δ

)

It should be noted that the conventional conﬁdence

of equation 1 may not ensure the desired inequality

(Eq. 5) as it depends only on the model initialisation

and the loss behavior during training.

FVC

≤ FVC

(5)

4.2 Model Optimization

The mean pinball loss function is used to train the pre-

dictive model as a quantile regression model. As de-

ﬁned in (Eq. 6), the value of the loss is equivalent to

half of the mean absolute error when the quantile pa-

rameter α is set to 0.5.

Pinball(y, ˆy) =

∑

α max(y

− ˆy

, 0)

+(1 −α) max( ˆy

−y

, 0) (6)

4.3 Evaluation Metric

Since the model should predict both the FVC and its

conﬁdence, we used the Laplace log likelihood (LLL)

as deﬁned in (Eq. 7). It is a well designed metric to

handle the uncertainty on the predictions.

To avoid penalization from large errors, the maxi-

mum error is limited to 1000 ml. The minimum con-

ﬁdence values are clipped to 70 as to reﬂect the ap-

proximate measurement uncertainty in FVC. The ﬁ-

nal score is calculated by averaging the metric across

all test set Patient-Weeks. For details and more spe-

ciﬁc considerations, refer to the modiﬁed version of

the LLL used in the challenge

clipped

= max(σ, 70)

∆ = min(



FVC

−FVC

pred



, 1000) (7)

score = −

√

2∆

clipped

− log

√

2σ

clipped

5 EXPERIMENTATION

The multimodal model portrayed in ﬁgure 9, aggre-

gates the clinical data-based model model

and the

CT image-based model model

, in what follows we

will refer to the multimodal model as model

CL,CT

The CT scan-based architecture uses efﬁcientNet-

B5 as a backbone model with an input channel dimen-

sion of 1 as the scans are single channel. The classiﬁer

layer of the backbone model was replaced by a linear

layer to extract the image embedding vector that is

concatenated with the clinical embeddings.

https://www.kaggle.com/competitions/osic-

pulmonary-ﬁbrosis-progression/overview/evaluation

Enhanced Multimodal Timely Prediction of Pulmonary Fibrosis Progression with Uncertainty Estimation from Chest CT Images and

Clinical Metadata

465

Figure 9: The multimodal model

CL,CT

concatenating the image model and the clinical data-based model

. It combines

three inputs, 10 continues, 2 categorical variables, and CT images.

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

466

In Figure 10, the predictions of the multiple quan-

tile regression models model

and model

CL,CT

are

shown using the validation dataset. The plots corre-

sponding to the low and high quantile, respectively,

FVC

@.2

and FVC

@.8

, suitably delimit the validation

data. The plot referring to FVC

@.5

is well aligned

with the groundtruth FVC data.

Figure 11 depicts the distribution of the FVC

prediction uncertainty of model

and model

CL,CT

on the validation data. The distributions show how

well the distribution is preserved for both the clinical

data-based model

and the scan-based model

CL,CT

Globally, the distributions are almost similar.

(a) Clinical data model

(b) Multimodal model

CL,CT

Figure 10: Multiple quantile regression: FVC - predictions

on the validation set.

The ﬁnal FVC values and the corresponding un-

certainty was formulated as a blending of model

and model

CL,CT

predictions (Eq. 8). The weighting

parameters w

CL,CT

= 0.1 and w

= 0.9 that repre-

sent the contribution of the respective model were ex-

perimentally determined.

FVC

Blend

= w

CL,CT

FVC

CL,CT

+ w

FVC

(8)

Table 2 summarizes the performance of the

blended model on the private test set. The baseline

model achieved a LLL score of -6.8272. Compared

to the LLL of the state of the art approach which ob-

(a) Clinical data model

(b) Multimodal model

CL,CT

Figure 11: Uncertainty distribution from validation data.

tained a score of -6.8305, our enhanced model

Blend

achieved a superior LLL of -6.8227 using the pro-

posed anti-Elu of the intermediate block and the anti-

Relu conﬁdence block.

Table 2: The Laplace log likelihood scores of the state-of-

the-art models.

Method LLL score ↓

Baseline (Lhagiimn, 2023) -6.8272

OSIC 1

place (Osic, 2023) -6.8305

- 2

place -6.8311

- 3

place -6.8336

model

(ours) -6.8234

model

Blend

(ours) -6.8227

6 CONCLUSION

In this article, we investigated the problem of pre-

dicting the progression of pulmonary ﬁbrosis (FVC)

from clinical data and CT-scan images. The prob-

lem was approached as a multiple quantile regression

problem. The visual scan data modality was com-

pared to clinical metadata modality, and then used to-

gether in a multimodal model. The introduced anti-

Enhanced Multimodal Timely Prediction of Pulmonary Fibrosis Progression with Uncertainty Estimation from Chest CT Images and

Clinical Metadata

467

Elu block intermediate block with the anti-Relu con-

ﬁdence block enhanced the multimodal timely predic-

tion of the pulmonary ﬁbrosis progression with uncer-

tainty estimation. The achieved performance against

the state-of-the-art models proved the effectiveness

of the proposed multimodal quantile regression-based

approach. This demonstrates once again that integrat-

ing the visual modality along with clinical metadata is

beneﬁcial for the robustness of the predictive model.

ACKNOWLEDGEMENTS

We gratefully acknowledge the support of the Com-

puter Research Institute of Montreal (CRIM) and the

Minist

ere de l’

Economie et de l’Innovation (MEI) of

Quebec.

REFERENCES

Agarwala, S. et al. (2020). Deep learning for screening of

interstitial lung disease patterns in high-resolution CT

images. Clinical Radiology, 75(6):481.e1–481.e8.

Christe, A. et al. (2019). Computer-aided diagnosis of pul-

monary ﬁbrosis using deep learning and CT images.

Investigative Radiology, 54(10):627–632.

Comelli, A. et al. (2020). Lung segmentation on high-

resolution computerized tomography images using

deep learning: A preliminary step for radiomics stud-

ies. Journal of Imaging, 6(11).

du Bois, R. M. et al. (2011). Forced vital capacity in pa-

tients with idiopathic pulmonary ﬁbrosis. American

Journal of Respiratory and Critical Care Medicine,

184(12):1382–1389.

Hutchinson, J. et al. (2015). Global incidence and mortality

of idiopathic pulmonary ﬁbrosis: a systematic review.

European Respiratory Journal, 46(3):795–806.

Kido, S. et al. (2022). Segmentation of Lung Nodules on

CT Images Using a Nested Three-Dimensional Fully

Connected Convolutional Network. Frontiers in Arti-

ﬁcial Intelligence, 5.

Lhagiimn (2023). www.kaggle.com/code/lhagiimn/solution-

for-the-ﬁrst-place-but-we-didn-t-select (accessed:

15.10.2023).

Mandal, S., Balas, V. E., Shaw, R. N., and Ghosh, A. (2020).

Prediction analysis of idiopathic pulmonary ﬁbrosis

progression from osic dataset. In 2020 IEEE Inter-

national Conference on Computing, Power and Com-

munication Technologies (GUCON), pages 861–865.

Nalysnyk, L., Cid-Ruzafa, J., Rotella, P., and Esser, D.

(2012). Incidence and prevalence of idiopathic pul-

monary ﬁbrosis: review of the literature. European

Respiratory Review, 21(126):355–361.

Noth, I. et al. (2021). Home spirometry in patients with id-

iopathic pulmonary ﬁbrosis: data from the INMARK

trial. Eur. Respir. J., 58(1):1–10.

Osic (2023). OSICcompetition leaderboard.

www.kaggle.com/competitions/osic-pulmonary-

ﬁbrosis-progression/leaderboard (accessed:

15.10.2023).

Raghu, G., Chen, S.-Y., Yeh, W.-S., Maroni, B., Li, Q.,

Lee, Y.-C., and Collard, H. R. (2014). Idiopathic

pulmonary ﬁbrosis in us medicare beneﬁciaries aged

65 years and older: incidence, prevalence, and sur-

vival, 2001–11. The Lancet Respiratory Medicine,

2(7):566–572.

Raghu, G. et al. (2006). Incidence and prevalence of idio-

pathic pulmonary ﬁbrosis. American Journal of Res-

piratory and Critical Care Medicine, 174(7):810–816.

Scatarige, J. C. et al. (2003). Utility of high-resolution ct

for management of diffuse lung disease: Results of a

survey of u.s. pulmonary physicians. Academic Radi-

ology, 10(2):167–175.

Soffer, S. et al. (2022). Artiﬁcial intelligence for interstitial

lung disease analysis on chest computed tomography:

A systematic review. Academic Radiology, 29:S226–

S235. Special Issue: Pulmonary.

Walsh, S. L. F. et al. (2018). Deep learning for classify-

ing ﬁbrotic lung disease on high-resolution computed

tomography: a case-cohort study. The Lancet Respi-

ratory Medicine, 6(11):837–845.

Watters, L. C. et al. (1986). A clinical, radiographic, and

physiologic scoring system for the longitudinal as-

sessment of patients with idiopathic pulmonary ﬁ-

brosis. American Review of Respiratory Disease,

133(1):97–103.

Wells, A. U. and Kokosi, M. A. (2016). Subclinical in-

terstitial lung abnormalities: Toward the early detec-

tion of idiopathic pulmonary ﬁbrosis? American

Journal of Respiratory and Critical Care Medicine,

194(12):1445–1446.

Wong, A. et al. (2021). Fibrosis-net: A tailored deep convo-

lutional neural network design for prediction of pul-

monary ﬁbrosis progression from chest CT images.

Frontiers in Artiﬁcial Intelligence, 4.

Yadav, A. et al. (2022). FVC-NET: An automated diagno-

sis of pulmonary ﬁbrosis progression prediction using

honeycombing and deep learning. Computational In-

telligence and Neuroscience, 2022:1–12.

Zucker, E. J. et al. (2020). Deep learning to automate

brasﬁeld chest radiographic scoring for cystic ﬁbrosis.

Journal of Cystic Fibrosis, 19(1):131–138.

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

468