Explainable AI Approach for Cardiac Involvement Detection in

Anderson-Fabry Disease

Chiara Verdone

1 a

, Matteo Gravina

2 b

, Grazia Casavecchia

2 c

, Rodolfo Belﬁore

and Benedetta Di Millo

Department of Agricultural Science, Food, Natural Resources and Engineering, University of Foggia, Foggia, Italy

Department of Medical and Surgical Sciences, Section of Diagnostic Imaging, University of Foggia, Foggia, Italy

ﬁ

Keywords:

Deep Learning, Healthcare, Anderson-Fabry Disease, Explainability.

Abstract:

Anderson-Fabry Disease (AFD) is a rare X-linked hereditary disorder caused by a deﬁciency of the enzyme

alpha-galactosidase A, leading to the accumulation of globotriaosylceramide (Gb3) in multiple organs, in-

cluding kidneys and the cardiovascular system. This study explores the role of deep learning techniques in the

analysis of cardiac imaging data for the early detection and monitoring of AFD-related cardiac involvement.

Using advanced image processing algorithms, we aim to improve diagnostic accuracy, assess myocardial ﬁ-

brosis progression, and facilitate personalized patient management. Our ﬁndings highlight the potential of

artiﬁcial intelligence in enhancing diagnostic workﬂows, reducing variability in interpretation, and aiding

clinicians in making more informed decisions. Furthermore, the use of non-invasive imaging techniques and

Native T1 sequences for mapping studies in cardiac magnetic resonance imaging (CMR) could reduce the

need for contrast.

1 INTRODUCTION

Anderson-Fabry disease (AFD) is an X-linked hered-

itary disease caused by total or partial deﬁciency

of the enzyme alpha-galactosidase A which leads

to the lysosomial accumulation of globotriaosylce-

ramide (Gb3) in many organs, including the kid-

neys, nervous system, cardiovascular system and skin

(Torra, 2008; Tøndel et al., 2008).

Among the affected organs, the heart is partic-

ularly vulnerable, with Gb3 deposition in myocytes

(heart muscle cells) and ﬁbroblasts leading to progres-

sive myocardial ﬁbrosis, cardiomyopathy (enlarge-

ment and weakening of the heart), ventricular dys-

function, and arrhythmias. Early and accurate cardiac

involvement detection is crucial, as myocardial ﬁbro-

sis signiﬁcantly worsens the prognosis of AFD pa-

tients. Cardiac Magnetic Resonance Imaging (CMR)

has become the non-invasive gold standard for evalu-

ating cardiac manifestations, particularly through the

use of Native T1 sequences (Deva et al., 2016).

Native T1 sequences provides valuable informa-

tion on myocardial tissue characteristics without the

https://orcid.org/0000-0003-1335-5276

https://orcid.org/0000-0001-9601-8253

https://orcid.org/0009-0000-9320-7833

need for contrast agents. This is particularly impor-

tant in AFD patients, many of whom may have renal

impairment, where the use of gadolinium-based con-

trast agents poses a risk of nephrogenic systemic ﬁ-

brosis and other complications.

Despite the effectiveness of CMR, the manual in-

terpretation of images remains subjective and prone

to variability. Artiﬁcial Intelligence (AI), particu-

larly Deep Learning techniques such as Convolutional

Neural Networks (CNNs), offers the potential to auto-

mate and standardize the evaluation process, improv-

ing diagnostic accuracy and reproducibility. How-

ever, to ensure clinical adoption, AI models must also

be explainable, enabling clinicians to understand the

rationale behind predictions (Aversano et al., 2024;

Denaro et al., 2024; Aversano et al., 2023).

In this work, we propose a CNN-based ap-

proach for the automated analysis of Native T1 car-

diac MRI scans to detect cardiac involvement in

AFD. Our pipeline integrates advanced preprocess-

ing techniques, model optimisation, and explainabil-

ity through Grad-CAM visualisation. We aim to

demonstrate that AI-driven, non-contrast-based diag-

nosis can support clinicians in managing AFD pa-

tients, reducing risks associated with contrast admin-

istration while maintaining high diagnostic reliability.

Verdone, C., Gravina, M., Casavecchia, G., Belﬁore, R., Di Millo and B.

Explainable AI Approach for Cardiac Involvement Detection in Anderson-Fabry Disease.

DOI: 10.5220/0013653700003967

In Proceedings of the 14th International Conference on Data Science, Technology and Applications (DATA 2025), pages 811-818

ISBN: 978-989-758-758-0; ISSN: 2184-285X

811

2 BACKGROUND

2.1 Cardiac Magnetic Resonance

Cardiac magnetic resonance imaging (CMR) with

Native T1 sequences is an advanced, non-invasive

method for the early diagnosis of lysosomal accumu-

lations in the heart, a crucial aspect of Fabry disease

(Deva et al., 2016). This approach allows the de-

tection of changes in the myocardium that may not

be visible with other diagnostic techniques, such as

echocardiography or electrocardiography. Key ben-

eﬁts of this approach are early detection, that is the

ability to identify changes in cardiac tissue before ob-

vious clinical symptoms appear allows for early inter-

vention, potentially slowing or stopping the progres-

sion of the disease. It’s important for continuous mon-

itoring because offers the ability to follow the evolu-

tion of myocardial ﬁbrosis and other cardiac changes

over time, which is essential for assessing the effec-

tiveness of treatments and the heart’s response to ther-

apy. The longitudinal relaxation time (T1) in native

sequences, such as the T1 Native used for tissue map-

ping studies, can vary depending on the magnetic res-

onance imaging (MRI) machine used. This depends

on several factors, including ﬁeld strength, sequences

and image parameters, tissue type, machine calibra-

tion an customization.

2.2 Convolutional Neural Networks in

Medical Imaging

The approach adopted in this study leverages Con-

volutional Neural Networks (CNNs), a class of deep

learning models extensively utilized in medical imag-

ing (Yamashita et al., 2018). CNNs consist of con-

volutional layers, pooling layers, and fully connected

layers. Convolutional layers extract features from in-

put images by applying ﬁlters that detect spatial pat-

terns, such as edges and textures. Pooling layers re-

duce the spatial dimensions, preserving essential in-

formation while improving computational efﬁciency.

Fully connected layers aggregate these features and

produce the ﬁnal classiﬁcation output.

CNNs have demonstrated remarkable success in

detecting and classifying diseases in medical imaging,

including cardiovascular and renal pathologies (Lit-

jens et al., 2017). However, their use in distinguishing

patients with concomitant cardiac and renal diseases

remains an area of active research.

2.3 Explainability in Deep Learning:

The Role of Grad-CAM

One of the main challenges of AI in healthcare is the

black-box nature of deep learning models. Clinicians

require transparency and interpretability to trust AI-

generated diagnoses. In this work, we employ the

Gradient-weighted Class Activation Mapping (Grad-

CAM) technique (Selvaraju et al., 2017) to enhance

model explainability.

Grad-CAM extends the Class Activation Mapping

(CAM) method (Zhou et al., 2016) by generating lo-

calization maps that highlight relevant regions in an

image for classiﬁcation. It computes the gradient of

the target class score with respect to the feature maps

of the last convolutional layer. These gradients are

then aggregated to assign importance weights to fea-

ture maps, producing heatmaps that overlay the origi-

nal MRI scans. This visualization helps clinicians un-

derstand which areas of the image contributed most

to the model’s decision, fostering trust in AI-assisted

diagnostics.

The integration of CNNs and explainability tech-

niques like Grad-CAM is essential for advancing AI-

driven diagnosis in cardiac and renal diseases. By

providing interpretable visualizations, this approach

bridges the gap between deep learning and clinical

applicability, ensuring that AI models support rather

than replace human expertise.

3 RELATED WORKS

In recent years, deep learning techniques have demon-

strated signiﬁcant potential in medical imaging analy-

sis, particularly in the classiﬁcation and segmentation

of cardiovascular and renal pathologies.

Several studies have explored the application of

Convolutional Neural Networks (CNNs) for the auto-

matic diagnosis of heart diseases using cardiac MRI.

For instance, (Bai et al., 2018) employed deep learn-

ing models to extract features from cardiac MRI

scans, achieving high accuracy in detecting various

cardiac abnormalities. Similarly, (Chen et al., 2019)

demonstrated that CNN-based architectures could ef-

fectively classify myocardial diseases, highlighting

the importance of spatial feature extraction.

Regarding kidney disease detection, recent works

have focused on integrating deep learning with MRI

scans to improve early diagnosis (Zhang et al.,

2024). For example, (Almuayqil et al., 2024)] uti-

lized a CNN-based approach to identify renal dis-

eases, achieving promising results in distinguishing

pathological cases from healthy ones. Additionally,

DMDH 2025 - Special Session on Data-Driven Models for Digital Health Transformation

812

Blur Image Removal

a. Image Preprocessor

b. Model Construction c. Explainability

Dark Image Removal Resize

288 x 288

128 x 128

Talos optimization

Prediction

CNN Model

GradCam heatmap

288 x 288

Data Augumentation

Input RM images

Best Model

Figure 1: Approach.

multi-modal AI models that combine MRI with other

clinical data have been proposed to enhance diagnos-

tic performance (Qiao et al., 2025). (Gravina et al.,

2024) introduced a non-invasive technique for identi-

fying renal complications in Anderson-Fabry disease

using advanced imaging methods.

Given the growing interest in explainable AI

(XAI), some research efforts have also focused on

developing interpretable models for medical imaging.

Attention mechanisms and Grad-CAM visualizations

have been employed to highlight the most relevant re-

gions in MRI scans, ensuring that AI-driven decisions

align with expert knowledge (Selvaraju et al., 2020).

In this context,(Prencipe et al., 2023) proposed an ex-

plainable radiogenomic framework to predict muta-

tional statuses in lung adenocarcinoma, demonstrat-

ing how AI can be both predictive and interpretable

in complex medical diagnoses.

Our study builds upon these advancements by de-

veloping a CNN-based approach for distinguishing

patients with both cardiac and renal diseases from

those with only cardiac conditions. By leveraging

deep learning techniques we aim to improve diagnos-

tic accuracy and interpretability in the classiﬁcation

of complex multi-organ pathologies.

4 APPROACH

Figure 1 shows the general approach adopted and we

observe that it is composed of 4 different phases: (a)

Image Preprocessor, (b) Model Construction and (c)

Explainability Task. The tasks will be detailed in

more detail in the following paragraphs.

4.1 Image Processor

The ﬁrst task initially focuses on extracting the tho-

racic magnetic resonance slices (in DICOM format)

from each study belonging to each patient. Subse-

quently, the extracted images were subjected to sev-

eral preprocessing steps to eliminate from the analy-

sis images that did not have an acceptable quality and

which could, therefore, introduce noise into the model

training process (Figure 1 a).

The ﬁrst preprocessing activity concerns the re-

moval of blurry images. To do this, a function writ-

ten in Python was used that uses the principle of the

variance of the Laplacian; it is a ﬁlter that measures

the variations in intensity of pixels in an image. In

particular, if an image has sharp and deﬁned details,

the Laplacian will have very variable values (high

variance); if, instead, an image is blurry or grainy,

the changes between pixels will be smaller (low vari-

ance). In this speciﬁc analysis, a double threshold ap-

proach was adopted: a lower one equal to 5 to check

if the image was blurry and an upper one equal to 400

to check if the image was noisy or grainy. The reasons

behind the choice of this technique are that ﬁrst of all,

the Laplacian is more effective on grayscale images as

in the case in question; it does not depend on the reso-

lution or orientation of the image, and furthermore, it

is used in radiology for image enhancement and noise

removal in medical scans (Fotin et al., 2019).

Explainable AI Approach for Cardiac Involvement Detection in Anderson-Fabry Disease

813

conv block 1

convolutional + ReLU

max pooling

fully connected + ReLU

softmax

input tensor

normalization

dropout

flatten

input shape

(128x 128)

conv block 2

conv block 4

binary output

conv block 3

conv block 5

Figure 2: CNN Architecture.

The second preprocessing step involves removing

images that are too dark compared to the average us-

ing an approach based on the image histogram. In

particular, it measures the proportion of pixels with

low-intensity values and compares it with a prede-

ﬁned threshold. The basic idea is that in dark images,

most pixels have low light intensity, so a high number

of pixels with values close to zero indicates a possible

underexposure. Also, in this case, a function written

in Python was used. Speciﬁcally, ﬁrst, all pixels with

an intensity value lower than 30 are counted, and then

the percentage of dark pixels compared to the entire

image is calculated; if this percentage is greater than

a threshold value set at 65%, the image is considered

too dark and is therefore discarded.

This technique was chosen because it does not de-

pend on the content of the image itself, and is also al-

ready used in the ﬁeld of medical imaging (Agarwal

and Mahajan, 2018).

To improve the model’s generalisation ability and

reduce the risk of overﬁtting, a data augmentation

strategy was implemented using the imgaug library

The applied transformations were chosen to sim-

ulate realistic variations in chest MRI images without

signiﬁcantly altering the anatomical features.

The augmentation operations adopted include:

• Slight rotation (-5° to 5°): useful to compensate

https://imgaug.readthedocs.io/en/latest/ last visit:

April 2025

for slight variations in the orientation of the scans.

• Scaling (0.95x – 1.05x): introduce slight changes

in the image size to improve the robustness of the

model at different acquisition resolutions.

• Random crop up to 5%: simulates variations in

the acquisition ﬁeld of view and helps the model

to focus on discriminant features.

• Contrast variation (0.4x – 1.6x): emulates differ-

ences in imaging conditions and acquisition pa-

rameters.

• Gaussian blur (sigma = 0 – 0.1) with probability

50%: introduce slight noise to improve the robust-

ness of the model to variations in image quality.

• Horizontal and vertical translations: to introduce

slight variations in position (3%) since MRIs may

be centred slightly differently between patients.

Finally, the image was resized to a size of 128 ×

128 pixels.

4.2 Model Construction

The model used to perform the classiﬁcation is based

on a 2-dimensional convolutional network (Figure 1

b) since it is a type of network widely used in medical

imaging (Anwar et al., 2018).

In particular, the model is made up of 5 con-

volutional layers interspersed with Normalization,

Dropout and Maxpooling layers.

DMDH 2025 - Special Session on Data-Driven Models for Digital Health Transformation

814

Table 1: Hyperparameters and values considered.

Hyperparameter Description Evaluated Values

Neurons Dimensionality of the output space; number

of output ﬁlters in the convolution.

128, 64, 32

Number of epochs Number of epochs to train the model. 10, 20

Batch size Number of samples per gradient update. 8, 16, 32, 64

Optimizer Algorithm used to decrease loss by tuning pa-

rameters and weights for better accuracy.

adam

Loss Function used to evaluate a candidate solu-

tion.

Binary Cross-Entropy

Dropout rate Probability of dropping a node during an iter-

ation.

0.15, 0.2

Last activation Activation function of the last layer. Sigmoid

In the Figure 2, the layers that compose each con-

volutional block are visible. In particular, we can no-

tice that each block is made up of a convolutional

layer followed by a normalization layer and a dropout

layer; after the ﬁrst, third and ﬁfth blocks, we ﬁnd a

pooling layer.

Thus, the model was subjected to optimising the

hyperparameters that characterize it using Talos tech-

nology

. Talos automates hyperparameter tuning in

deep learning models, allowing researchers and de-

velopers to explore various hyperparameter combina-

tions using search methods such as grid search and

random search. Talos lets you easily set up a series of

hyperparameter experiments, run them, and analyze

them to ﬁnd the best conﬁguration.

In addition to the convolutional network built

speciﬁcally for the analysis, a pre-trained VGG16

(provided by Keras

) network was also tested to eval-

uate the performance of a state-of-the-art model using

the dataset in question.

4.3 Explainability Task

The explainability task is built on the application of

the GradCam technique. Both the CNN model and

the predictions are sent as input. GradCam generates

a heat map to locate the regions most inﬂuencing the

model predictions in chest MRI scans related to Fabry

disease.

As shown in the Figure 1 c represents the Grad-

CAM map for a diseased subject. In particular, we can

see the region that signiﬁcantly inﬂuences the classiﬁ-

cation of Fabry disease, which is highlighted in green.

https://autonomio.github.io/talos last visit: April 2025

https://keras.io last visit: April 2025

5 EXPERIMENTAL

DESCRIPTION

The adopted dataset and experiment settings will be

described in the following paragraphs.

5.1 Dataset

To validate the constructed model, a real dataset was

used, whose data were collected at the Azienda Os-

pedaliero Universitaria Policlinico Riuniti di Foggia,

Italy.

The data are related to 31 cardiac patients divided

into 17 men with an average age of approximately

40.65 years and 14 women with an average age of

47.79 years.

Of these 31 patients, 13 have cardiac involvement

in Fabry disease, so they are labelled as positive in the

analysis, while 18 patients do not have a correlation,

so they are marked as negative.

Some patients had multiple visits over the years,

so they were considered multiple times. Considering

that each scan can have a variable number of slices in-

side it, downstream of the slice extraction phase, we

have a total of 1621 images, from which 260 images

were removed because they were blurry and 28 be-

cause they were dark; therefore, the resulting dataset

is composed of 1333 grayscale images with a size of

288 x 288 pixels.

As for the data augmentation task, starting from

each image another one was generated using one of

the techniques, randomly chosen, mentioned in the

previous paragraph 4.1. In this way, the ﬁnal dataset

used for the analysis is composed of 2666 images.

5.2 Experiment Setting

The experiments aim to evaluate the performance of

the proposed approach.

Explainable AI Approach for Cardiac Involvement Detection in Anderson-Fabry Disease

815

Table 2: Results obtained with hyperparameters optimization and ablative study.

Parameters Metrics Ablative Study

Low Quality

Network Neurons Epochs Batch size Dropout Accuracy Precision Recall F-score Data Augumentation Image Balancing

Removal

CNN 64 20 8 0.15 0.82 0.8082 0.8254 0.8129 yes no no

CNN 32 20 8 0.15 0.8117 0.8003 0.7913 0.7952 yes yes no

VGG16 128 20 8 0.2 0.7194 0.7787 0.7194 0.7037 yes yes yes

CNN 128 10 64 0.15 0.8917 0.9021 0.8917 0.891 yes yes yes

The classiﬁcation process involves splitting each

considered sub-dataset into training, validation, and

test sets. Speciﬁcally, 30% of the dataset is allocated

for testing, while the remaining 70% is further divided

into 70% for training and 30% for validation.

The experiment is structured as an ablative study.

The ﬁrst analysis concerned the use of the CNN net-

work to apply to the set of images resulting from the

data augmentation operation.

The second analysis adds the component related

to the removal of low-quality images, so after having

removed them, we proceeded to apply the data aug-

mentation operation; the network used in this case is

the same CNN network used before.

Considering the strong imbalance between the

classes, it was decided to apply a subsampling tech-

nique of the majority class so that the model sees the

same number of images for each class without syn-

thetically creating new samples. For this reason, the

third analysis adds this balancing component to the

data, always using the same CNN network.

This same experiment conﬁguration was also

tested with the VGG16 network to obtain a compari-

son with a pre-trained network.

Regarding performance evaluation, the goal is to

assess the efﬁciency of the proposed classiﬁcation

process on the considered sub-datasets. The evalua-

tion is conducted using the following metrics: Accu-

racy, Precision, Recall, and F1-score, which are de-

rived from the Confusion Matrix.

Additionally, a hyperparameter optimization step

is performed using Talos, which enables testing all

possible combinations of hyperparameters in a single

experiment. The best model is selected and saved for

future predictions.

The Table 1 outlines the considered hyperparame-

ters and their evaluated values.

The training phase for each classiﬁer was done

twice, and the model with the highest average F1-

score was chosen. In the event of models with iden-

tical F1-scores, the one with the greatest average Ac-

curacy was chosen.

Finally, the experiment attempts to assess the sug-

gested approach’s capacity to deliver right interpreta-

tions using Grad-CAM heatmaps manual reviewed by

a team of specialists.

6 DISCUSSION OF RESULTS

This section shows and discusses the obtained results.

Table 2 reports the results obtained using the CNN

and VGG16 networks after undergoing hyperparame-

ter optimization using Talos.

In particular, the ﬁrst column indicates the type of

network used, columns 2 to 5 indicate the hyperpa-

rameters that make up the best conﬁguration for the

network being analyzed, the next 4 columns indicate

the validation metrics, and ﬁnally, the last 3 columns

indicate the ﬂags related to each task of the ablative

study.

As we can see, the best result is obtained using

all three tasks for the ablative study and the custom

CNN network, reaching an accuracy of 89.1% and a

precision of 90.2%.

In order to further explore the differences in the

behaviour of the two networks considered, the confu-

sion matrices of both are reported relative to the last

conﬁguration of the experiment.

In particular, we note that in the case of the

VGG16 network (Table 3), this presents a large num-

ber of correct predictions for patients with AFD with

cardiac involvement, but on the contrary, it misclassi-

ﬁes many of the patients who do not present cardiac

involvement, leading to signiﬁcant consequences if it

were used as a decision-making tool for the physician.

A different situation occurs for the confusion ma-

trix related to the CNN network (Table 4); ﬁrst of

all we can notice that the main diagonal is the most

populated symptom of the fact that the prediction was

much more correct than the previous case, as can also

be seen from the validation metrics calculated starting

from the matrix (Table 2, last row).

We also notice only 5 misclassiﬁcations regarding

patients who do not present cardiac involvement; we

can, therefore, conclude that the CNN network is able

to predict the pathology in a more correct and bal-

anced way.

DMDH 2025 - Special Session on Data-Driven Models for Digital Health Transformation

816

Table 3: VGG16 Confusion Matrix.

AFD HC

AFD 171 9

HC 92 88

Table 4: CNN Confusion Matrix.

AFD HC

AFD 146 34

HC 5 175

In addition to the quantitative measurement of the

model performance, in this study, we also report a

qualitative evaluation analysing the obtained interpre-

tation maps (heat maps) and the feedback of an expert.

A measure, therefore, to better understand where

the model has concentrated most to discriminate the

classes.

Figure 3: Grad-CAM heatmap for a patient with AFD and

cardiac implications.

In Figure 3, we observe an example of a heatmap

for a patient affected by AFD with cardiac involve-

ment obtained by applying the Grad-CAM technique

associated with the best model. In particular, the pix-

els in yellow represent the pixels that assume greater

importance in the model, and we can note that in ad-

dition to highlighting the kidney at the bottom, they

also highlight the cardiac area, indicating cardiac in-

volvement in the AFD pathology. We can therefore

conclude that the model is focusing on the right area

of the image to make its decisions.

7 CONCLUSIONS

The implementation of deep learning-based method-

ologies in the assessment of cardiac involvement in

Anderson-Fabry Disease has shown promising re-

sults. Our study demonstrates that AI-driven analysis

of cardiac imaging can improve diagnostic precision

and contribute to earlier detection of myocardial ﬁ-

brosis. By reducing subjectivity and enhancing repro-

ducibility in image interpretation, these approaches

can help clinicians optimize treatment strategies. The

scope of creating a radiomic model for Native T1

assessment is ambitious, but it can certainly lead to

signiﬁcant improvements in patient management, re-

ducing gadolinium dependency and providing useful

information for the diagnosis and monitoring of car-

diac diseases. However, it will be essential to de-

velop a robust pipeline that includes accurate segmen-

tation, feature extraction, and clinical validation to en-

sure that the model can effectively replace LGE se-

quences. Additionally, the ability to detect cardiac

abnormalities using non-invasive imaging methods,

such as CMR, could help avoid unnecessary contrast-

enhanced CT scans, improving patient safety and

comfort. Future research should focus on integrating

multi-modal imaging data and reﬁning AI models to

further enhance their clinical applicability. The con-

tinued evolution of AI in medical imaging holds great

potential for improving patient outcomes in rare ge-

netic diseases like AFD.

REFERENCES

Agarwal, M. and Mahajan, R. (2018). Medical image con-

trast enhancement using range limited weighted his-

togram equalization. Procedia Computer Science,

125:149–156.

Almuayqil, S. N., Abd El-Ghany, S., Abd El-Aziz, A. A.,

and Elmogy, M. (2024). Kidneynet: A novel

cnn-based technique for the automated diagnosis of

chronic kidney diseases from ct scans. Electronics,

13(24).

Anwar, S. M., Majid, M., Qayyum, A., Awais, M., Al-

nowami, M., and Khan, M. K. (2018). Medical im-

age analysis using convolutional neural networks: a

review. Journal of medical systems, 42:1–13.

Aversano, L., Bernardi, M. L., Cimitile, M., Iammarino, M.,

Madau, A., and Verdone, C. (2023). An explainable

approach for early parkinson disease detection using

deep learning. Communications in Computer and In-

formation Science, 1875 CCIS:326 – 339. Cited by:

Aversano, L., Bernardi, M. L., Cimitile, M., Montano, D.,

Pecori, R., and Veltri, L. (2024). Explainable anomaly

detection of synthetic medical iot trafﬁc using ma-

chine learning. SN Computer Science, 5(5). Cited

by: 3; All Open Access, Green Open Access, Hybrid

Gold Open Access.

Bai, W., Sinclair, M., Tarroni, G., Oktay, O., Rajchl, M.,

Vaillant, G., Lee, A. M., Aung, N., Lukaschuk, E.,

Sanghvi, M. M., et al. (2018). Automated cardiovas-

cular magnetic resonance image analysis with fully

Explainable AI Approach for Cardiac Involvement Detection in Anderson-Fabry Disease

817

convolutional networks. Journal of cardiovascular

magnetic resonance, 20(1):65.

Chen, M., Fang, L., Zhuang, Q., and Liu, H. (2019). Deep

learning assessment of myocardial infarction from mr

image sequences. IEEE Access, 7:5438–5446.

Denaro, F., Madau, A., Martini, C., and Pecori, R. (2024).

An explainable approach to characterize heart dis-

eases using ecg images. In 2024 IEEE International

Conference on Metrology for eXtended Reality, Arti-

ﬁcial Intelligence and Neural Engineering, MetroX-

RAINE 2024 - Proceedings, page 867 – 872. Cited

by: 0.

Deva, D. P., Hanneman, K., Li, Q., Ng, M. Y., Wasim, S.,

Morel, C., Iwanochko, R. M., Thavendiranathan, P.,

and Crean, A. M. (2016). Cardiovascular magnetic

resonance demonstration of the spectrum of morpho-

logical phenotypes and patterns of myocardial scar-

ring in anderson-fabry disease. Journal of Cardiovas-

cular Magnetic Resonance, 18(1):14.

Fotin, S. V., Yankelevitz, D. F., Henschke, C. I., and

Reeves, A. P. (2019). A multiscale laplacian of gaus-

sian (log) ﬁltering approach to pulmonary nodule de-

tection from whole-lung ct scans. arXiv preprint

arXiv:1907.08328.

Gravina, M., Troise, D., Infante, B., Tartaglia, L., Minop-

oli, B., Allegra, C., Casavecchia, G., Gambacorta,

M., Montanile, C., Mercuri, S., et al. (2024). A

non-invasive technique to unveil renal implications in

anderson–fabry disease. Biomedicines, 12(9):1950.

Litjens, G., Kooi, T., Bejnordi, B. E., Setio, A. A. A.,

Ciompi, F., Ghafoorian, M., van der Laak, J. A., van

Ginneken, B., and S

anchez, C. I. (2017). A survey

on deep learning in medical image analysis. Medical

Image Analysis, 42:60–88.

Prencipe, B., Delprete, C., Garolla, E., Corallo, F., Gravina,

M., Natalicchio, M. I., Buongiorno, D., Bevilacqua,

V., Altini, N., and Brunetti, A. (2023). An explainable

radiogenomic framework to predict mutational status

of kras and egfr in lung adenocarcinoma patients. Bio-

engineering, 10(7):747.

Qiao, Y., Zhou, H., Liu, Y., Chen, R., Zhang, X., Nie,

S., Hou, F. F., Zhao, Y., Xu, X., and Zhao, L.

(2025). A multi-modal fusion model with enhanced

feature representation for chronic kidney disease pro-

gression prediction. Brieﬁngs in Bioinformatics,

26(1):bbaf003.

Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R.,

Parikh, D., and Batra, D. (2017). Grad-cam: Visual

explanations from deep networks via gradient-based

localization. In Proceedings of the IEEE international

conference on computer vision, pages 618–626.

Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R.,

Parikh, D., and Batra, D. (2020). Grad-cam: visual

explanations from deep networks via gradient-based

localization. International journal of computer vision,

128:336–359.

Tøndel, C., Bostad, L., Hirth, A., and Svarstad, E. (2008).

Renal biopsy ﬁndings in children and adolescents with

fabry disease and minimal albuminuria. American

journal of kidney diseases, 51(5):767–776.

Torra, R. (2008). Treatment and renal involvement in fabry

disease. Kidney International, 74(Suppl 111):S29–

S32.

Yamashita, R., Nishio, M., Do, R. K. G., and Togashi, K.

(2018). Convolutional neural networks: an overview

and application in radiology. Insights into imaging,

9:611–629.

Zhang, M., Ye, Z., Yuan, E., Lv, X., Zhang, Y., Tan,

Y., Xia, C., Tang, J., Huang, J., and Li, Z. (2024).

Imaging-based deep learning in kidney diseases: re-

cent progress and future prospects. Insights into imag-

ing, 15(1):50.

Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., and Tor-

ralba, A. (2016). Learning deep features for discrim-

inative localization. In Proceedings of the IEEE con-

ference on computer vision and pattern recognition,

pages 2921–2929.

DMDH 2025 - Special Session on Data-Driven Models for Digital Health Transformation

818