Explainability Applied to a Deep-Learning Based Algorithm for Lung

Nodule Segmentation

Arman Zafaranchi

1,2

, Francesca Lizzi

, Alessandra Retico

, Camilla Scapicchio

1,2

and

Maria Evelina Fantacci

1,2

National Institute for Nuclear Physics (INFN), Pisa, Italy

Department of Physics, University of Pisa, Pisa, Italy

Keywords: Lung Nodule Detection, Lung Segmentation, Deep Learning, Segmentation, Lung Cancer.

Abstract: Deep learning and computer-aided detection (CAD) methods play a pivotal role in the early detection and

diagnosis of various cancer types. The significance of AI in the medical field has become particularly

pronounced during the coronavirus pandemic. This study aims to develop a deep learning-based system for

segmenting and detecting nodules in the lung parenchyma, utilizing the Luna-16 challenge dataset. The

algorithm is divided into two phases: the first phase involves lung segmentation using the previously

developed LungQuant algorithm to identify the region of interest (ROI), and the second phase employs a

specifically designed and fine-tuned Attention Res-UNet for nodule segmentation. Additionally, the

explainable AI (XAI) technique, Grad-CAM, was used to demonstrate the reliability of the proposed

algorithm for clinical application. In the initial phase, the LungQuant algorithm achieved an average Dice

Similarity Coefficient (DSC) of 90%. For nodule segmentation, the DSC scores were 81% test sets. The model

also achieved average sensitivity and specificity metrics of 0.86 and 0.92.

1 INTRODUCTION

Lung cancer imposes a significant global health

burden, with an alarming annual incidence of over 1.6

million new cases worldwide. As the second most

common form of cancer, it surpassed breast cancer in

incidence among women in developed nations.

Despite advances in medical technology, the

prognosis for lung cancer remains challenging

(Houda et al., 2024).

Early detection of lung cancer is crucial for

effective treatment and improved survival rates

(Mohamed et al., 2024). Despite physical symptoms

(Durstenfeld et al., 2022), more accurate diagnostic

methods are necessary to initiate treatment. Computed

Tomography (CT) is a highly sensitive imaging

modality. However, frequent CT scans, as required by

possible screening programs, can lead to overexposure

to ionizing radiation. To mitigate this risk, Low Dose

CT (LDCT) scans are now employed for high-risk

patients, allowing the reduction of radiation exposure

through advanced reconstruction and analysis software

(Barca et al., 2018). LDCT is effective in detecting

early-stage lung cancer and has been shown to reduce

mortality rates by 20% (Silva et al., 2022).

Medical image analysis is a challenging task that

requires a high degree of concentration and substantial

expertise, with significant variability among

specialists. This is particularly true in the context of

lung cancer, where small nodules indicate positive

cases, yet these nodules frequently lack uniform size,

volume or location which make them difficult to

detect. This variability is crucial during the early stages

of treatment and can greatly affect a patient's long-term

survival prospects (Peters et al., 2021).

The significance of AI in medical imaging has been

further underscored during the COVID-19 pandemic,

where researchers have developed CAD systems to aid

in detecting infected lesions in lung CT scans. These

AI-powered tools serve as invaluable aids to

radiologists, enhancing diagnostic accuracy and

expediting patient care processes (Greenspan et al.,

2020).

During the COVID-19 pandemic, researchers

developed several CAD systems (Karimkhani et al.,

2022; Lizzi et al., 2023) to assist physicians in

detecting infected lesions in lung CT scans. AI-based

software has proven to be a supportive tool for

radiologists, capable of highlighting potential

abnormalities in CT scans that might be overlooked,

132

Zafaranchi, A., Lizzi, F., Retico, A., Scapicchio, C. and Fantacci, M.

Explainability Applied to a Deep-Learning Based Algorithm for Lung Nodule Segmentation.

DOI: 10.5220/0013014600003886

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

In Proceedings of the 1st International Conference on Explainable AI for Neural and Symbolic Methods (EXPLAINS 2024), pages 132-138

ISBN: 978-989-758-720-7

thereby prompting further review or additional tests

by human experts. (Gozes et al., n.d.) developed a

deep learning-based CT image analysis system that

could accurately differentiate between COVID-19

positive and negative patients. This system localized

lung abnormalities and provided quantitative

measurements, supporting radiologists' diagnostic

and prognostic assessments.

The AI system consisted of multiple components,

analysing CT cases at two levels: 3D analysis for

nodules and focal opacities using existing algorithms,

and 2D analysis of each slice to detect larger diffuse

opacities, such as ground-glass infiltrates.

Additionally, (Fang et al., 2021) designed an AI-

powered framework to assess disease severity and

predict outcomes for COVID-19 patients. This

framework was evaluated using datasets from two

hospitals and compared against manual assessments

by radiologists, demonstrating superior accuracy in

predicting ICU admissions and mortality. The study

highlighted the potential of AI-based methodologies

to enhance the management of COVID-19 patients

(Scapicchio et al., n.d.).

The AI system's performance was compared to eight

human observers and the clinical assessments of

patients, including RT-PCR testing. The findings

revealed that CORADS-AI successfully automated

the scoring of chest CT scans, aligning with the CO-

RADS and CT severity score metrics, and performed

comparably to human observers in terms of CT

severity scores, with equal or superior proficiency in

identifying COVID-19 positive patients.

In recent years, deep learning (DL) methods have

emerged as powerful tools for medical image

analysis, offering significant improvements in the

segmentation of lung nodules. These methods

leverage large datasets and complex algorithms to

identify and delineate nodules with high precision.

One such algorithm, adapted from the LungQuant

approach, forms the foundation of our method’s

initial phase in finding the ROI.

Despite their potential, the "black-box" nature of DL

models raises concerns about their transparency and

interpretability, which are crucial for clinical

adoption. Therefore, incorporating XAI techniques is

imperative to ensure the transparency and reliability

of these models, thereby fostering trust among

medical professionals. We will present our approach

to lung nodule segmentation using DL methods,

supplemented by XAI results, to demonstrate the

accuracy and interpretability of our models. By doing

so, we aim to highlight the transformative potential of

DL in lung cancer diagnosis and advocate for the

integration of XAI in clinical practice.

2 MATERIAL AND METHODS

The main goal of our project is to create a reliable and

robust CAD for lung cancer detection utilizing deep

learning methods. In the first step of our paper we

implemented a two-step algorithm using the Luna-16

dataset alongside with explainable AI techniques to

demonstrate the reliability of the model. Fig.1

illustrates the schematic representation of the

proposed algorithm.

2.1 Dataset

A noteworthy dataset used in our study is the Lung

Nodule Analysis 2016 challenge (Luna-16) (Murphy

et al., 2009), renowned for its application in lung

cancer detection. Comprising CT scans from 888

patients, Luna-16 provides ground truth information

for ROI segmentation, along with the coordinates of

nodules in a 3D scale. Luna-16 is derived from the

LIDC-IDRI dataset, featuring specific nodule

volumes and low-dose CT screening. For the first

phase obviously, we used original CTs with ground

truth of lung parenchyma for training. Then, for the

second phase of the algorithm, we generated a 3D

cube with nodules in the determined coordinates, so

during the training process each slice of segmented

ROI can match with the generated mask. Before

segmentation, we applied initial preprocessing to the

CT scans, which included normalizing the image

intensities and the Hounsfield Unit of CTs.

2.2 Phase 1: Lung Segmentation

Lung nodule segmentation is a challenging task for

AI due to factors such as image noise, imbalanced

data, and the complex structure of lung tissues. To

address these challenges, we implemented several

techniques. In the initial step, identifying the ROI

helps to reduce the complexity of the image structure.

For this purpose, we utilized first part of the

LungQuant algorithm to segment the lung region

from body organs in CT scans.

LungQuant is a fully automated deep learning-

based system designed to assist radiologists in

detecting lung lesions indicative of COVID-19

infection (Lizzi et al., 2022). The initial version,

introduced in 2023, demonstrated significant

promise. A subsequent version was released with a

refined structure to enhance the segmentation

accuracy of lung parenchyma and COVID-19

pneumonia in CT scans (Lizzi et al., 2023). This

section will explore the details of the LungQuant

methodology.

Explainability Applied to a Deep-Learning Based Algorithm for Lung Nodule Segmentation

133

Figure 1: Diagram of proposed algorithm.

LungQuant was developed using deep learning

algorithms in multiple steps and have been evaluated

to asses with various datasets (Scapicchio et al.,

2023). Initially, an AlexNet-based DNN predicts two

points to define a bounding box around the 3D voxel

data of the lungs, aiding in the localization of the lung

parenchyma for further analysis. The next phase

employs two U-nets: the first segments the lung

parenchyma, which we have utilized in this paper,

and the second uses these results to accurately

identify and delineate COVID-19 lesions. Pre-

processing and data augmentation were applied to

prevent overfitting and improve model performance.

2.3 CLAHE Preprocessing

After segmenting the lung region using LungQuant,

we applied Contrast Limited Adaptive Histogram

Equalization (CLAHE) to the segmented lung images

(Kyriakopoulou, 2020). CLAHE is an advanced

image preprocessing technique used to enhance the

contrast of images, particularly in medical imaging

for improving the visibility of features within an

image. This technique improves the contrast of an

image in a localized manner, making it easier to

detect features like lung nodules in medical images.

By limiting the contrast enhancement, CLAHE

reduces the risk of noise amplification while

preserving fine details and edges in the image, which

is crucial for accurate diagnosis and analysis in

medical imaging.

2.4 Phase 2: Nodule Segmentation

In the second phase of our methodology, we focus on

the segmentation of lung nodules using an advanced

deep learning model. This phase builds upon the

output of the first phase, where the lung region was

isolated using the LungQuant algorithm.

To achieve accurate nodule segmentation, we

employed an Attention Res-UNet architecture. This

model is designed to enhance the focus on relevant

features while maintaining the spatial details crucial

for precise nodule detection. The Attention Res-UNet

incorporates attention blocks that selectively

highlight important features in the image, reducing

the impact of irrelevant background information. This

mechanism improves the model’s ability to detect

small and subtle nodules amidst the lung parenchyma.

Moreover, the architecture utilizes residual

connections, allowing the model to learn more

effectively by mitigating the vanishing gradient

problem. This enhancement helps in preserving the

gradient flow through deep layers, ensuring better

learning of complex patterns. For the training process,

we generated 3D cubes with nodules at the specified

coordinates provided by the LUNA-16 dataset. Using

the nodule coordinates from the dataset, we created

binary masks for each nodule. These masks are

essential for training the model, providing the ground

truth for the nodule locations. We used the Dice Loss

function (Sudre, C.H., Li, W., Vercauteren, T.,

Ourselin, S., Jorge Cardoso, 2017), which is

particularly effective for imbalanced data, where

background voxels are more than nodules one. The

fine-tuning process involved training on the

generated data and refining the model’s architecture

to enhance its ability to distinguish nodules from

surrounding tissue, thereby yielding promising results

in lung nodule segmentation This pre-processing

ensures that each slice of the segmented ROI can be

matched with the corresponding mask.

EXPLAINS 2024 - 1st International Conference on Explainable AI for Neural and Symbolic Methods

134

Figure 2: Results of Lung segmentation with LungQuant’s first phase.

2.5 Model Explanation and

Performance Evaluation

Explainable AI is crucial in various applications,

especially in high-stakes fields like healthcare, for

building trust and transparency in order to demystify

the “black box” nature of deep learning models to

make their decision transparent. Moreover, In

healthcare, decisions based on AI can have significant

consequences. XAI ensures that AI models can be

held accountable for their decisions, providing

explanations that can be analysed.

To ensure the interpretability of our model, we

applied the Grad-CAM (Gradient-weighted Class

Activation Mapping) technique. Grad-CAM

(Selvaraju et al., 2016) is a powerful visualization

tool that helps in understanding and interpreting the

decisions made by deep learning models. It highlights

the regions in the input image that contribute most

significantly to the model's predictions, thereby

providing a visual explanation of the model's focus

and attention. For each CT scan slice processed by the

Attention Res-UNet, we generated Grad-CAM

heatmaps.

These heatmaps were overlaid on the original CT

images to highlight the regions where the model

focused its attention while identifying nodules. The

visual explanations provided by Grad-CAM helped in

validating the model’s predictions by confirming

whether the identified regions correspond to actual

nodules. This step is crucial for gaining the trust of

medical professionals and ensuring the reliability of

the AI system. By analysing the Grad-CAM

heatmaps, we could identify any potential areas

where the model might be making incorrect

predictions or missing nodules. This feedback loop

allowed us to fine-tune the model and improve its

performance iteratively.

2.6 Metrics

To assess the performance of each phase, we applied

appropriate metrics for thorough evaluation and

comparison. For the first phase, lung segmentation

performance validation, we used the DSC to measure

the overlap between prediction and ground truth. For

the second phase of nodule segmentation, we utilized

sensitivity, specificity, and the average False Positive

Rate (FPR) per scan. These metrics provide a

comprehensive evaluation of the algorithm's accuracy

and reliability in both lung region segmentation and

nodule detection.

3 RESULTS

In this section, we present the outcomes of our study

on lung nodule segmentation using DL methods,

supported by XAI. The results are organized to

demonstrate the efficacy of our approach, the

performance of the model, and the interpretability of

its decisions.

Up to this point, we have elaborated on the details

of the proposed algorithm. Broadly speaking, we have

three distinct objectives in this paper. The first

objective is to use and evaluate the performance of

LungQuant for lung segmentation purposes. By

achieving this, we aim to obtain a more precise ROI

and demonstrate the robustness of our deep learning-

based algorithm. This step is crucial in ensuring the

accuracy of subsequent phases and in showcasing the

efficacy of LungQuant in clinical applications. In the

original LungQuant paper, a 96% DSC was achieved

on the COVID-19-CT-Seg dataset. For our first

objective, we evaluated the lung segmentation task

using DSC and obtained an average score of 90%

based on the provided ground truth. Fig. 2

demonstrates the algorithm's robustness across

different datasets and highlights LungQuant's

exceptional performance in the more challenging

regions of the lung, i.e. the bases.

In the second phase, we developed an Attention

Res-UNet architecture specifically for the nodule

segmentation task. To enhance the clarity of lung

tissue and reduce noise, we applied CLAHE to the

outputs from the first step. This preprocessing step

Explainability Applied to a Deep-Learning Based Algorithm for Lung Nodule Segmentation

135

Figure 3: Prediction of Attention Res-Unet.

Figure 4: Results of Grad-Cam for Explainability of Nodule segmentation.

was essential for improving the visibility of subtle

features within the lung images. Subsequently, we

fine-tuned the Attention Res-UNet model, optimizing

its parameters to achieve robust performance in

detecting and segmenting lung nodules. The trained

neural network achieved Dice Coefficients of 85%,

83%, and 81% for the training, validation, and test

sets, respectively. Additionally, the model reached

average sensitivity and specificity metrics of 0.86 and

0.92, with an average FPR of 2.25 per scan,

demonstrating its effectiveness and reliability in lung

nodule segmentation. Figure 3 showcases the

accurate segmentation results of our fine-tuned model

for nodule detection. The comparison between the

predicted points and the generated mask highlights

the model’s outstanding performance.

Final objective of this paper is to visualize the

areas where the Attention Res-UNet model focused

during prediction. Grad-CAM generates heatmaps

that highlight important regions in the input image for

predicting lung nodules, providing insights into the

model’s decision-making process. In Fig. 4, the Grad-

CAM visualization shows a focused heatmap around

a small, distinct region within the lung parenchyma.

The highlighted region corresponds to a suspected

nodule, indicating that the model successfully

identified this area as important for nodule detection.

The concentration of the heatmap around the nodule

demonstrates the model's ability to localize the

nodule accurately. Moreover, in the case with

presence of two nodules the high-intensity heatmap

accurately highlights the nodule's location.

4 DISCUSSION

As mentioned before, our project’s goal is to develop

a deep learning-based CAD algorithm for lung cancer

detection. Up to this point, we have designed, fine-

tuned, and tested several complex deep neural

networks to evaluate and compare the performance of

different models, i.e. U-Net, Res U-Net, Attention U-

Net, on the LUNA-16 dataset.

Recent research indicates that attention

mechanisms can perform well with complex data like

medical images. Specifically, in our scenario of

detecting lung nodules with low volume amidst lung

tissues, the attention mechanism can effectively focus

on the target parts. Additionally, residual blocks help

to mitigate the vanishing gradient issue, which is

likely due to the similar structure of the data.

One of the long-term goals of this project is to

implement the developed algorithm in clinical

environments, which necessitates ensuring the

reliability and robustness of the CAD system. The

integration of our proposed DL-based methodology,

particularly the use of LungQuant for lung

segmentation, and an Attention Res-UNet for nodule

EXPLAINS 2024 - 1st International Conference on Explainable AI for Neural and Symbolic Methods

136

segmentation, has the potential to improve diagnostic

workflows in clinical settings. This approach can

assist radiologists by providing accurate and reliable

segmentation, thereby reducing workload and

improving early detection rates of lung cancer.

Incorporating XAI techniques, such as Grad-

CAM, is vital for guaranteeing the transparency and

trustworthiness of AI models in medical imaging.

XAI offers insights into the model’s decision-making

process, thereby enhancing the interpretability and

acceptance of AI-based tools by medical

professionals.

In this process, we encounter several challenges.

One limitation of our study is the relatively small

dataset size, which may impact the generalizability

and robustness of our results. Furthermore, variations

in image quality and the assumptions made during

model training and evaluation could influence the

overall performance. To handle some of these issues

for our future research we intend to focus on

expanding the dataset to include more diverse cases,

further improving the model architecture, and

integrating additional preprocessing techniques to

enhance segmentation accuracy. Moreover, extensive

clinical trials are necessary to validate the efficacy of

the proposed methodology in real-world clinical

environments.

5 CONCLUSIONS

In this study, we emphasize the critical role of deep

learning-based CAD systems in the detection of lung

cancer using CT datasets, highlighting the importance

of early detection in improving patient survival rates.

We employed the LungQuant automated system for

segmenting the lung region and demonstrated the

generalization of this algorithm with different

datasets, achieving an average of 90% DSC with

Luna-16, in comparison to the 96% reported in the

original study. We then applied CLAHE

preprocessing to reduce noise and enhance tissue

details in the lung parenchyma. These pre-processed

images were input into an Attention Res-UNet for the

nodule segmentation task, resulting in DSC scores of

85%, 83%, and 81% for the training, validation, and

test sets, respectively. The model achieved average

sensitivity and specificity metrics of 0.86 and 0.92,

with an average FPR of 2.25 per scan. Our findings

indicate that attention mechanisms and residual

blocks significantly enhance segmentation

performance, even in complex scenarios. This work

underscores the transformative potential of deep

learning and explainable AI in lung cancer diagnosis,

advocating for their integration into clinical practice

to improve patient outcomes. For future work, we aim

to further refine the model to reduce the false positive

rate per scan, thereby enhancing its clinical utility and

reliability.

ACKNOWLEDGEMENTS

Research partly supported by: Artificial Intelligence

in Medicine (next AIM, https://www.pi.infn.it/aim)

project, INFN-CSN5; PNRR - M4C2 -

Partenariato Esteso ”FAIR - Future Artificial

Intelligence Research” - Spoke 8, funded by the

European Commission under the NextGeneration

EU programme; the European Union

NextGenerationEU through the Italian Ministry

of University and Research under PNRR M4C2-I1.3

Project PE 00000019 ”HEAL ITALIA” to Maria

Evelina Fantacci and Arman Zafaranchi CUP

I53C22001440006; European Union -

NextGenerationEU through the Italian Ministry of

University and Research under PNRR - M4C2-I1.5 -

Project ECS00000017 “Tuscany Health Ecosystem

(THE)” - CUP I53C21000350006.

The views and opinions expressed are those of the

authors only and do not necessarily reflect those of

the European Union or the European Commission

Neither the European Union nor the European

Commission can be held responsible for them.

REFERENCES

Barca, P., Palmas, F., Fantacci, M. E., & Caramella, D.

(2018). Evaluation of the adaptive statistical iterative

reconstruction algorithm in chest CT (Computed

Tomography) a preliminary study toward its

employment in low dose applications, also in

conjunction with CAD (Computer Aided Detection).

HEALTHINF 2018 - 11th International Conference on

Health Informatics, Proceedings; Part of 11th

International Joint Conference on Biomedical

Engineering Systems and Technologies, BIOSTEC

2018, 5, 688–694. https://doi.org/10.5220/

0006750706880694

Durstenfeld, M. S., Sun, K., Tahir, P., Peluso, M. J., Deeks,

S. G., Aras, M. A., Grandis, D. J., Long, C. S., Beatty,

A., & Hsue, P. Y. (2022). Use of Cardiopulmonary

Exercise Testing to Evaluate Long COVID-19

Symptoms in Adults: A Systematic Review and Meta-

analysis. In JAMA Network Open (Vol. 5, Issue 10, pp.

E2236057–E2236057). American Medical

Association. https://doi.org/10.1001/jamanetworkopen.

2022.36057

Explainability Applied to a Deep-Learning Based Algorithm for Lung Nodule Segmentation

137

Fang, X., Kruger, U., Homayounieh, F., Chao, H., Zhang,

J., Digumarthy, S. R., Arru, C. D., Kalra, M. K., & Yan,

P. (2021). Association of AI quantified COVID-19

chest CT and patient outcome. International Journal of

Computer Assisted Radiology and Surgery, 16(3), 435–

445. https://doi.org/10.1007/s11548-020-02299-5

Gozes, O., Frid-Adar, M., Greenspan, H., Browning, P. D.,

Zhang, H., Ji, W., ... & Siegel, E. (2020). Rapid ai

development cycle for the coronavirus (covid-19)

pandemic: Initial results for automated detection &

patient monitoring using deep learning ct image

analysis. arXiv preprint arXiv:2003.05037.Greenspan,

H., San José Estépar, R., Niessen, W. J., Siegel, E., &

Nielsen, M. (2020). Position paper on COVID-19

imaging and AI: From the clinical needs and

technological challenges to initial AI solutions at the lab

and national level towards a new era for AI in

healthcare. Medical Image Analysis, 66.

https://doi.org/10.1016/j.media.2020.101800

Houda, I., Dickhoff, C., Uyl-de Groot, C. A., Reguart, N.,

Provencio, M., Levy, A., Dziadziuszko, R., Pompili, C.,

Di Maio, M., Thomas, M., Brunelli, A., Popat, S.,

Senan, S., & Bahce, I. (2024). New systemic treatment

paradigms in resectable non-small cell lung cancer and

variations in patient access across Europe. The Lancet

Regional Health - Europe, 38, 100840.

https://doi.org/10.1016/j.lanepe.2024.100840

Karimkhani, H., Attariabad, A., & Vahed, H. (2022). High

sensitive plasmonic sensor with simple design of the

ring and the disk resonators. Optical and Quantum

Electronics, 54(6). https://doi.org/10.1007/s11082-

022-03736-2

Kyriakopoulou, M. (2020). Histogram Equalization on

Medical Images: CLAHE implementation on CT

images.

Lizzi, F., Agosti, A., Brero, F., Cabini, R. F., Fantacci, M.

E., Figini, S., Lascialfari, A., Laruina, F., Oliva, P.,

Piffer, S., Postuma, I., Rinaldi, L., Talamonti, C., &

Retico, A. (2022). Quantification of pulmonary

involvement in COVID-19 pneumonia by means of a

cascade of two U-nets: training and assessment on

multiple datasets using different annotation criteria.

International Journal of Computer Assisted Radiology

and Surgery, 17(2), 229–237.

https://doi.org/10.1007/s11548-021-02501-2

Lizzi, F., Postuma, I., Brero, F., Cabini, R. F., Fantacci, M.

E., Lascialfari, A., Oliva, P., Rinaldi, L., & Retico, A.

(2023). Quantification of pulmonary involvement in

COVID-19 pneumonia: an upgrade of the LungQuant

software for lung CT segmentation. European Physical

Journal Plus, 138(4). https://doi.org/10.1140/epjp/

s13360-023-03896-4

Mohamed, E., García Martínez, D. J., Hosseini, M. S.,

Yoong, S. Q., Fletcher, D., Hart, S., & Guinn, B. A.

(2024). Identification of biomarkers for the early

detection of non-small cell lung cancer: a systematic

review and meta-analysis. Carcinogenesis, 45(1–2), 1–

22. https://doi.org/10.1093/carcin/bgad091

Murphy, K., van Ginneken, B., Schilham, A. M. R., de

Hoop, B. J., Gietema, H. A., & Prokop, M. (2009). A

large-scale evaluation of automatic pulmonary nodule

detection in chest CT using local image features and k-

nearest-neighbour classification. Medical Image

Analysis, 13(5), 757–770. https://doi.org/10.1016/j.

media.2009.07.001

Peters, A. A., Decasper, A., Munz, J., Klaus, J., Loebelenz,

L. I., Hoffner, M. K. M., Hourscht, C., Heverhagen, J.

T., Christe, A., & Ebner, L. (2021). Performance of an

AI based CAD system in solid lung nodule detection on

chest phantom radiographs compared to radiology

residents and fellow radiologists. Journal of Thoracic

Disease, 13(5), 2728–2737. https://doi.org/10.

21037/jtd-20-3522

Scapicchio, C., Ballante, E., Brero, F., Cabini, R. F.,

Chincarini, A., Fantacci, M. E., ... & Retico, A. (2023).

Integration of a Deep Learning-Based Module for the

Quantification of Imaging Features into the Filling-in

Process of the Radiological Structured Report.

In HEALTHINF (pp. 663-670).

Scapicchio, C., Chincarini, A., Ballante, E., Berta, L., Bicci,

E., Bortolotto, C., Brero, F., Cabini, R. F., Cristofalo,

G., Fanni, S. C., Fantacci, M. E., Figini, S., Galia, M.,

Gemma, P., Grassedonio, E., Lascialfari, A., Lenardi,

C., Lionetti, A., Lizzi, F., … Retico, A. (2023). A

multicenter evaluation of a deep learning software

(LungQuant) for lung parenchyma characterization in

COVID-19 pneumonia. European Radiology

Experimental, 7(1). https://doi.org/10.1186/s41747-

023-00334-z

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

Parikh, D., & Batra, D. (2016). Grad-cam: Why did you

say that? visual explanations from deep networks via

gradient-based localization. Grad-CAM: Visual

Explanations from Deep Networks via Gradient-Based

Localization, 17, 331–336.

http://arxiv.org/abs/1610.02391

Silva, M., Picozzi, G., Sverzellati, N., Anglesio, S.,

Bartolucci, M., Cavigli, E., Deliperi, A., Falchini, M.,

Falaschi, F., Ghio, D., Gollini, P., Larici, A. R.,

Marchianò, A. V, Palmucci, S., Preda, L., Romei, C.,

Tessa, C., Rampinelli, C., & Mascalchi, M. (2022).

Low-dose CT for lung cancer screening: position paper

from the Italian college of thoracic radiology.

Radiologia Medica, 127(5), 543–559.

https://doi.org/10.1007/s11547-022-01471-y

Sudre, C.H., Li, W., Vercauteren, T., Ourselin, S., Jorge

Cardoso, M. (2017). (2017). Generalised Dice Overlap

as a Deep Learning Loss Function for Highly

Unbalanced Segmentations. Springer, vol 10553(Deep

Learning in Medical Image Analysis and Multimodal

Learning for Clinical Decision Support).

EXPLAINS 2024 - 1st International Conference on Explainable AI for Neural and Symbolic Methods

138