YOLOv8-Based Lung Nodule Detection: A Novel Hybrid Deep Learning Model Proposal

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 10 Issue: 08 | Aug 2023 www.irjet.net p-ISSN: 2395-0072
© 2023, IRJET | Impact Factor value: 8.226 | ISO 9001:2008 Certified Journal | Page 230
YOLOv8-Based Lung Nodule Detection: A Novel Hybrid Deep Learning
Model Proposal
CİHAT ŞAMAN, ŞERİFE ÇELİKBAŞ
1 BIOMEDİCAL ENGINEER , SMNWAY INC & cihatsaman@smnway.com, TURKEY
2LECTURER, İSTANBUL AYDIN UNIVERSITY & serifecelikbas@aydin.edu.tr, TURKEY
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Small oval or circular masses identified in the
lungs are known as lung nodules. When these nodules are
smaller than 3 cm, they are often considered less
concerning; however, over time, they may increase in size,
potentially leading to more serious consequences. Early
detection of these nodules and timely preventive measures
are crucial to impede their progression to malignancy.
Conventional diagnostic methods like computerized
tomography (CT) and radiographic imaging techniques are
utilized for this purpose. Nevertheless, these approaches can
either subject patients to excessive radiation or prove
inadequate in detecting small nodules. As a result, various
deep learning-based image processing techniques are being
explored for lung nodule detection. In this study, a novel
deep-learning model is proposed for the automated real-
time detection of lung nodules. The proposed model exhibits
a remarkable accuracy of 92.3% in nodule detection, along
with a sensitivity of 88.5% and a mean average precision
(mAP) of 53.5%. The model is built using the YOLOv8
architecture, with the YOLOv8m configuration yielding the
best results. Additionally, graphical comparisons with
existing studies in the literature demonstrate the
effectiveness of the training model.
Key Words: deep learning, lung cancer detection,
yolov8x, yolov8l, yolov8m, yolov8s, yolov8n
1. INTRODUCTION
Pulmonary nodules are small, round, or oval-shaped
growths in the lungs that can be benign or malignant. They
are often detected incidentally on chest imaging or
through dedicated screening programs for lung cancer.
The presence of pulmonary nodules may cause symptoms
such as cough, chest pain, shortness of breath, or coughing
up blood, especially if the nodules are large, multiple, or
cancerous. Some nodules may also be associated with
infections, inflammation, or autoimmune diseases that can
affect other organs and systems. Therefore, determining
the nature and cause of pulmonary nodules is essential for
providing appropriate treatment and preventing
complications. The epidemiology of pulmonary nodules
depends on various factors, such as the prevalence of
smoking, environmental exposures, infectious diseases,
and genetic susceptibility. According to the latest
guidelines and evidence, the management of pulmonary
nodules should be based on the risk of malignancy, which
can be estimated by using validated models that
incorporate radiographic and clinical features. The size,
shape, density, location, and growth rate of the nodules
are important radiographic characteristics that influence
risk assessment and follow-up recommendations. The use
of low-dose computed tomography (LDCT) for lung cancer
screening has increased the detection of pulmonary
nodules, especially those that are sub-solid or ground-
glass in appearance, which have a higher risk of cancer
than solid nodules. New criteria for the classification and
management of atypical pulmonary cysts, juxta pleural
nodules, and inflammatory or infectious findings have
been introduced in Lung-RADS® v2022, a standardized
reporting system for LDCT screening. The goal of
pulmonary nodule evaluation is to identify and treat lung
cancer at an early stage while minimizing unnecessary
interventions and harm for benign nodules [1-3].
Lung cancer is one of the most common and deadliest
types of cancer in the world [4]. Early diagnosis and
treatment of lung cancer is critical to improve survival. For
this purpose, developed computational image processing
studies are carried out. However, lung nodule detection
and diagnosis poses challenges for image processing and
analysis. Lung nodules can vary greatly in size, shape,
density, location, and number. In addition, factors such as
noise, artifact, and lack of contrast in images may
complicate the detection and characterization of lung
nodules. Lung nodule detection using the deep learning
method is an important step for early detection of lung
cancer. Deep learning techniques for lung nodule detection
provide higher performance and accuracy than traditional
computer-aided diagnostic systems. Zhang et al. [5]
utilized 3D DenseNet and 3D FPN models for detecting
lung nodules. They designed a dense feature pyramid
network to extract multi-scale features from different
layers of the DenseNet backbone. In their study, they
achieved an impressive competition performance metric
(CPM) value of 0.8934 on the LUNA16 dataset, showcasing
compelling results. Furthermore, the detection
performance was improved by approximately 2%
compared to other methods Zhang et al. [6] proposed an
integrated active contour model (IACM_MRFEBPD) for
small ground glass opacity (GGO) pulmonary nodule
segmentation. The method combines Markov random field
energy and Bayesian probability difference, achieving an
average IOU of 0.7444, 0.7503, and 0.7450 for LIDC-IDRI

test set, clinical test set, and all test sets, respectively. The
approach enables accurate and robust segmentation,
aiding in early lung cancer detection through medical
imaging evaluation. Ye et al. [7] proposed an automated
pulmonary nodule detection model using modified V-Nets
and an advanced SVM classifier. The model achieved
66.7% accuracy, 81.1% true positive rate, 78.1% positive
predictive value, and 78.7% F-score. Experimental results
on the LUNA16 dataset demonstrate its effectiveness in
early lung cancer detection. Huang et al. [8] propose a fast
and fully automated end-to-end system for precise
pulmonary nodule analysis. Their approach achieves high
accuracy, with 91.4% and 94.6% detection rates and an
average of 1 and 4 false positives (FPs) per scan. The
nodule segmentation exhibits an average dice coefficient
of 0.793, demonstrating its effectiveness in automated
lung nodule analysis. Luo et al. [9] present an anchor-free
3D sphere-based lung nodule detection network. The
approach automatically predicts nodule position, radius,
and offset without predefined anchor parameters. The
model effectively performs existing methods on the
LUNA16 dataset, achieving high true positive rates (TPR)
of 92.2 (1 FPs/image), 93.9 (2 FPs/image), and 96.4 (8
FPs/image) for lung nodule detection. Vipparla et al. [10]
propose three diverse 3D Attention-based CNN
architectures for lung nodule detection. The models use
attention mechanisms and contextual dependencies to
achieve significant improvements in classification
performance. Fusion of their predictions leads to more
effective results with a CPM score of 0.931 on the LUNA16
dataset. Han et al. [11] propose a CAD system, Pulmonary
Nodules Detection Assistant Platform, for early detection
and classification of pulmonary nodules in physical
examination LDCT images. Their 3D CNN-based model
achieves an accuracy of 0.879 with an average of 1 false
positive per CT in LNPE1000 dataset. The system shows
efficiency in detecting smaller and larger nodules, offering
potential benefits for early lung cancer detection in
physical examination applications. Alakwaa et al. [12]
propose a computer-aided diagnosis (CAD) system for
lung cancer classification using CT scans with unmarked
nodules. The system achieves an accuracy of 86.6% and
requires less labeled data, offering a more efficient and
generalized approach to lung cancer diagnosis.
Lung nodule detection was addressed in this study using a
two-stage deep learning model. In the first stage, image
processing filters were employed to enhance learning
performance and highlight nodule regions. Subsequently, a
Convolutional Neural Network (CNN) was trained,
achieving an impressive 94% accuracy. The resulting
weight model was then utilized to detect cancerous
regions. For the next step, the nodule regions identified by
the CNN and those marked by radiology experts were
prepared as input windows and trained with YOLOv8
models. The performance metrics, including accuracy
(92.3%), precision (87.3%), and mean Average Precision
(mAP - 53.5%), were obtained based on the training data.
The study results were presented in tables and compared
with findings from the existing literature.
The other sections of the paper are organized as follows:
Section 2 introduces the dataset and presents the
explanations of the methods and techniques used. Section
3 explains the general workflow of the proposed
convolutional model. Section 4 presents experimental
results and comparative analyses demonstrating the
effectiveness of the proposed approach. Finally, Section 5
includes evaluations related to the obtained results.
2. MATERIAL METOD
2.1. Dataset
The LUNA16 dataset is derived from the Lung Image
Database Consortium (LIDC-IDRI) database, which has
been publicly released. The LIDC-IDRI database contains
CT images and expert radiologists' annotations. The
LUNA16 dataset consists of 888 chest CT scans in DICOM
format. The dataset provides DICOM files along withs .mhd
and..zraw extensions. The annotations for nodule regions
were determined by expert radiologists. The dataset
includes annotations for both malignant (cancerous) and
benign (non-cancerous) nodules. Nodules vary in terms of
their sizes, shapes, and locations within the lungs. In this
study, the LUNA16 dataset was processed using the
Python-3.11.1 torch-2.0.1+cpu CPU (12th Gen Intel Core
(TM) i7-12700H) hardware in the PyCharm IDE, and a
deep learning model was proposed for nodule detection
[13].
2.2. Convolutional Neural Network
The increasing advancement and accessibility of hardware
technologies, coupled with the abundance of data in
various fields, have positively influenced the proliferation
and success of deep learning studies [15]. Convolutional
Neural Network (CNN) is a deep learning method that
processes its inputs by analyzing specific features through
different layers for classification tasks. Its ability to
determine features on its own provides ease of use. As a
result, it is effectively employed in various medical
applications such as cancer diagnosis and nodule
detection [16-18]. CNN, as shown in Figure 1, consists of a
convolutional layer, pooling layer, activation layer, and
fully connected layer. The convolutional layer works based
on the logic of applying filters of a predetermined size to
perform operations on the image and extract overlapping
regions. It demonstrates efficiency in detecting low-level
features such as edge detection, gradient orientation, and
color.

Fig -1: CNN architecture [19]
After the convolutional layer, the pooling layer contributes
to reducing the dimensionality of the data, thereby
reducing the computational power required to process it.
It also aids in extracting spatially invariant features and
helps prevent overfitting. An activation function is applied
to the computation results to finalize the output and
create a mathematical value. Subsequently, the obtained
outputs are passed to the fully connected layer to be
transformed into a single vector and then forwarded to
classification functions [19].
In this study, images obtained from .mhd extension DICOM
files were processed with the CNN architecture shown in
Figure 4 after the preprocessing steps. In addition to the
general layers, the dropout layer is also used in the
architecture. In order to increase the learning
performance, the learning rate was reduced after 10
epochs and satisfactory results were obtained. The regions
marked by the weight model formed as a result of the
training were recorded in .txt format. In these txt files, png
coordinates are retained and the corresponding
coordinates are determined when the coordinates shared
in the LUNA16 dataset are converted to 2D form. For each
png, the positions determined by both the CNN
architecture and the expert radiologists in the data set
were taken as the YOLO input. The results obtained are
presented in Chapter 4.
2.3. You Only Look Once (YOLO)
YOLO (You Only Look Once) is an object detection
algorithm that has become increasingly popular in recent
years, reducing the need for high-performance hardware.
YOLO is fundamentally a deep learning algorithm based on
CNN and is commonly used in applications such as
segmentation, classification, object detection, and real-
time object tracking [20, 21]. Its initial version was
published in 2015, and since then, various versions have
been developed to maintain a balance between speed and
high accuracy.
Real-time nodule detection in medical data is crucial for
achieving fast and high-performance results with deep
learning [13]. Therefore, this study utilizes YOLOv8 for
this purpose Accordingly, this study proposes a two-stage
learning system for medical data. In the first stage, an
abundant and sufficient amount of data is preprocessed
and trained with the CNN architecture. After the training
and nodule identification, in the second stage, YOLOv8 is
fine-tuned for real-time applications to achieve fast and
convenient integration.
YOLOv8 is created in 5 different scale types. It can learn
features such as object detection, segmentation, pose
estimation, tracking, as well as multiple views in different
dimensions. These scales; YOLOv8n (nano), YOLOv8s
(small), YOLOv8m (medium), YOLOv8l (large), and
YOLOv8x (extra-large) [22]. This study also presents the
results of these five different scales for the same epoch
comparatively.
3. PROPOSED MODEL
A hybrid model is being proposed to improve the
accuracy, efficiency, and real-time capability of lung
nodule detection and enhance the differentiation between
benign and malignant nodules. The overall workflow of
the study is illustrated in Figure 2.
Initially, the data from the LUNA16 dataset was
preprocessed to determine the lung parenchyma by
adjusting the values to the Hounsfield unit scale [-1000,
400][13]. Preprocessing steps were then applied to each
DICOM slice, followed by combining each slice with other
images from the corresponding .mhd extension file. This
process yielded a total of 888 .png files. These files were
subsequently divided into training, testing, and validation
sets using the CNN model, and then trained with the CNN
model, resulting in the creation of a weight file (.h5).
Using the obtained weight file, nodule detection was
performed on all 888 images, and the predicted regions
containing nodules were annotated. The coordinates of
these regions were recorded and compared with the
average coordinates provided by expert radiologists in the
LUNA16 dataset using an .mhd file. In cases where there

were significant differences in the coordinates, two sets of
coordinates were generated. Coordinates that were in
close proximity to each other were merged to form a
single rectangle, and the data was saved.
Fig -2: Workflow diagram of the proposed model
Subsequently, the acquired data and image inputs were
trained using YOLOv8, which offers fast training and real-
time processing capabilities. The learning outputs were
presented along with performance metrics. Finally, the
trained weight model was validated for each DICOM slice.
3.1. İmage Preprocess
After extracting the files with the .mhd extension from the
segmentation folder containing 888 CT images in the
LUNA16 dataset, the first step involved extracting the lung
parenchyma based on the Hounsfield Unit scale.
Subsequently, a series of image-processing steps were
applied to enhance the lung parenchyma and facilitate
subsequent analysis. These steps included min-max
normalization to ensure consistent data scaling and
prevent errors due to unusual data values. Furthermore,
the following procedures were performed: Unsharp
Masking, Median Filtering, Laplacian Filtering,
Binarization, Mask Negation, Morphological Closure, Mask
Negation (Repeated), Hollow Filling, and Mask
Subtraction. The accompanying Figure 3 illustrates the
progressive changes in the parenchymal image as each
procedure was applied.
Fig -3: Image preprocessing pipelines
Hounsfield Units are characterized as a quantitative
measurement of radio intensity. It is used to make regional
comments on CT and MR images and to analyze the
existing tissue type. For this study, HU scaling was
performed in the range of [-1000, 400]. These values were
chosen because they highlight the lung parenchyma tissue
[13]. After determining the pakanchyme texture, min-max
normalization was applied to scale the intensity values of
each pixel in the image to the [0, 1] range. The
normalization formula is as follows:
X' = (X - min_density) / (max_density - min_density) (1)
Here X represents the original pixel intensity value, X'
represents the normalized pixel intensity value,
min_density, and max_density represent the minimum and
maximum intensity values in the original range. In the
second stage, unsharp masking was applied. It is used to
highlight edges and details in the image by removing a
blurred version of the original image from the original
image. A median filter, which is a noise reduction
technique that replaces the pixel value with the median

value, has been applied in order to eliminate the noise
generated after this process. It is the Laplace filtering
technique, which is an edge detection technique that
emphasizes rapid density changes in the image in order to
more precisely detect the nodules in the image. After this
filtering step, a binarization process determined according
to a certain threshold was applied in order to analyze the
nodules more easily. And in the final stage, mask negation
process was applied in order to emphasize the nodules
more and reduce the noise from the background.
After the creation of the data, the images were adjusted to
512x512 for training purposes and passed through the
CNN model given in Figure 4. A convolutional layer is used,
which creates feature maps by scanning with various
filters and is effective in detecting edges, color changes
and other important patterns in the image. 32 3x3 filters
and relu layers as activation functions, 64 3x3 filters and
relu layers as activation functions are used, respectively.
In order to reduce the computational load and reduce the
risk of overfitting, maximum pooling is applied in size (2,
2). In order to provide input to the dense layer, it flattens
the output of Conv2D and MaxPooling2D layers into a
single vector. In the Dense layer, important features are
learned for classification by combining their features. In
the last stage, the SoftMax function is used to determine
the class prediction. The weight model with .h5 extension
obtained at the end of the process was recorded.
Fig -4: Proposed CNN architecture
After 150 epochs of training, we achieved a satisfactory
training success rate of 94%. Detailed training outcomes
are extensively documented in Chapter 4. The obtained
model was employed for detecting nodules, and each
nodule's boundary was enclosed within a rectangle. The
resulting data from this process were stored in separate
.txt files, which serve as input data for YOLOv8.
Additionally, the data information annotated by expert
radiologists in .mhd files was also incorporated into these
.txt files. Subsequently, we conducted training with 5000
epochs using the ultrasonic library on the yolov8.pt model.
The results obtained were highly satisfactory and
demonstrated comparability with findings reported in the
literature. Section IV provides a comprehensive account of
the experimental results.
4. EXPERİMENTAL RESULTS
In this study, the evaluation metrics employed for lung
nodule detection include accuracy, sensitivity, recall,
precision, mAP (mean Average Precision), and mAP50-95
(mean Average Precision between 50 and 95). The
obtained results are based on the outputs of TP (True
Positive), TN (True Negative), FP (False Positive), and FN
(False Negative) generated after the training process.
Accuracy represents the percentage of correctly classified
nodules relative to the total number of detected nodules. It
is calculated using the formula:
Accuracy = (TP + TN) / (TP + TN + FP + FN) (2)
Sensitivity measures the percentage of true positive cases
relative to the sum of true positive and false negative
cases. The formula is as follows:
Sensivity = TP / (TP + FN) (3)
Recall denotes the sensitivity percentage relative to the
sum of true positive and false negative cases. The formula
is given by:
Recall = TP / (TP + FN) (4)
Precision indicates the percentage of true positive cases
relative to the sum of true positive and false positive cases.
The formula is as follows:
Precision = TP / (TP + FP) (5)

mAP(mean Average Precision) represents the average
value of sensitivity at different recall levels. The formula is
given by:
mAP = (1 / N) * Σ(Recall) (6)
mAP50-95(mean Average Precision between 50 and 95
reflects the average sensitivity between recall levels of
50% to 95%. The formula is given by:
mAP50-95 = (1 / (95 - 50 + 1)) * Σ(Recall) (for i from 50 to
95) (7)
In this study, the primary objective is to enable real-time
and rapid detection of benign or malignant nodules within
the region captured by CT images. Lung nodule detection
is achieved by training a Convolutional Neural Network
(CNN) using information from nodule regions based on
radiologist annotations. The dataset consists of 888 CT
images in DICOM format, each containing corresponding
.mhd and .zraw files. Through image processing and
preprocessing steps, 551,065 PNG images were generated.
Notably, the 93rd and 94th PNG layers, which generally
display the lung parenchyma for each CT, were included in
the training process. The training achieved 94% accuracy
over 150 epochs.
Subsequently, an .h5 model was created after training.
Using this model, nodules were detected in the images and
saved to a .txt file. Additionally, the coordinates marked by
radiology experts in the candidates.csv file of the LUNA16
dataset were converted into 2D .png planes. The resulting
coordinate information was also saved in the relevant .txt
file. Both the generated .txt files and images were utilized
as input for YOLOv8 and evaluated for five different scales
over 30 epochs. The obtained results have been
summarized in Table 1. According to these results, when
evaluated at different scales, the YOLOv8x model
demonstrated the highest precision value. Furthermore,
the mAP (mean average precision) and mAP50-95 (mean
average precision at IoU thresholds of 50 to 95) values
showed the best performance for the YOLOv8m model.
These findings indicate that the YOLOv8 model achieved
high success in the object detection task, and particularly,
the YOLOv8x and YOLOv8m models outperformed others
when compared across different scales.
Simultaneously, these findings provide evidence that the
proposed methodology is suited and confirm its
practicality, particularly in the context of real-time
applications. The results underscore the methodology's
potential to excel when supplied with more
comprehensive training data and deployed on robust
hardware.
Table 1. Performance Evaluation of YOLOv8 Scales
Trained over 30 Epochs
YOLO
scales
Precision mAP50 mAP50-
95
Yolov8n 0.291 0.356 0.296
Yolov8s 0.775 0.505 0.394
Yolov8m 0.873 0.535 0.4
Yolov8l 0.626 0.523 0.303
Yolov8x 0.894 0.482 0.36
Moreover, the study results exhibit comparability with
similar previous works, as presented in Table 2,
showcasing the performance outcomes in a comparative
manner. The findings highlight the efficacy and potential
of the proposed approach for lung nodule detection,
contributing to the advancement of medical imaging and
early lung nodule diagnosis.
Table 2. Performance Evaluation of Studies Utilizing the
LUNA16 Dataset
Study Acc. Sens. Method Used
Ye et al.
[7]
66.7
%
81.1% Modified V-Nets and
Advanced SVM Classifier
Huang et
al. [8]
91.4
%
- Fast and Fully Automated
End-to-End System
Han et al.
[11]
87.9
%
- 3D CNN-Based Model for
Physical Examination LDCT
Images
Alakwaa
et al. [12]
86.6
%
- Computer-Aided Diagnosis
(CAD) System for Lung
Cancer Classification Using CT
Scans with Unmarked
Nodules
Propose
d
92.3
%
88.5
%
The model trained using 2D
Convolutional Neural
Networks (2DCNN) and
YOLOv8
5. CONCLUSIONS
A deep learning model is proposed in this research to
facilitate disease diagnosis with reduced radiation
exposure. The detection of nodules and cancerous tissues
is targeted using real-time imaging devices, such as C-arm
fluoroscopy, which involve lower radiation doses. A novel

deep learning model, specifically tailored for the detection
of lung nodules, enables real-time automated detection of
predefined nodules. Lung nodule detection is achieved
with an accuracy of 92.3%, a sensitivity of 88.5%, and a
mean average precision (mAP) of 53.5%. When compared
to the YOLOv8 architecture, superior results are obtained
with the YOLOv8m architecture. Furthermore, graphical
comparisons with related studies in the literature
effectively demonstrate the effectiveness of the training
model. This study represents a significant stride towards
early lung nodule detection and the implementation of
preventive measures, offering valuable insights for future
research. Deeper networks and larger datasets can be
explored in future studies to further enhance the
robustness and accuracy of the results.
REFERENCES
[1] Callister, M.E. et al. (2015) ‘British Thoracic Society
Guidelines for the investigation and management of
pulmonary nodules: Accredited by nice’, Thorax, 70(Suppl
2), pp. ii1–ii54. doi:10.1136/thoraxjnl-2015-207168.
[2] Larici, A.R. et al. (2017) ‘Lung nodules: Size still
matters’, European Respiratory Review, 26(146), p.
170025. doi:10.1183/16000617.0025-2017.
[3] Mazzone, P.J.; L. (2022) Evaluating the patient with a
pulmonary nodule, jama, DeepDyve. Available at:
https://www.deepdyve.com/lp/american-medical-
association/evaluating-the-patient-with-a-pulmonary-
nodule-Z1JwqOiXDs (Accessed: 06 August 2023).
[4] Chen, X., Mo, S. and Yi, B. (2022) ‘The spatiotemporal
dynamics of lung cancer: 30-year trends of epidemiology
across 204 countries and Territories’, BMC Public Health,
22(1). doi:10.1186/s12889-022-13281-y.
[5] Zhang, M. et al. (2021) ‘Pulmonary nodule detection
based on 3D feature Pyramid Network with incorporated
squeeze‐and‐excitation‐attention mechanism’,
Concurrency and Computation: Practice and Experience,
35(16). doi:10.1002/cpe.6237.
[6] Zhang, S. et al. (2020) ‘Segmentation of small ground
glass opacity pulmonary nodules based on Markov
random field energy and bayesian probability difference’,
BioMedical Engineering OnLine, 19(1).
doi:10.1186/s12938-020-00793-0.
[7] Ye, Y. et al. (2020) ‘Pulmonary nodule detection using
V-net and high-level descriptor based SVM classifier’, IEEE
Access, 8, pp. 176033–176041.
doi:10.1109/access.2020.3026168.
[8] Huang, X. et al. (2019) ‘Fast and fully-automated
detection and segmentation of pulmonary nodules in
thoracic CT scans using deep convolutional neural
networks’, Computerized Medical Imaging and Graphics,
74, pp. 25–36. doi:10.1016/j.compmedimag.2019.02.003.
[9 ] Luo, X. et al. (2022) ‘SCPM-net: An anchor-free 3D lung
nodule detection network using sphere representation
and center points matching’, Medical Image Analysis, 75, p.
102287. doi:10.1016/j.media.2021.102287.
[10] Vipparla, V.K., Chilukuri, P.K. and Kande, G.B. (2021)
‘Attention based multi-patched 3D-cnns with hybrid
fusion architecture for reducing false positives during lung
nodule detection’, Journal of Computer and
Communications, 09(04), pp. 1–26.
doi:10.4236/jcc.2021.94001.
[11] Han, Y. et al. (2022) ‘Pulmonary nodules detection
assistant platform: An effective computer aided system for
early pulmonary nodules detection in physical
examination’, Computer Methods and Programs in
Biomedicine, 217, p. 106680.
doi:10.1016/j.cmpb.2022.106680.
[12] Alakwaa, W., Nassef, M. and Badr, A. (2017) ‘Lung
cancer detection and classification with 3D Convolutional
Neural Network (3D-CNN)’, International Journal of
Advanced Computer Science and Applications, 8(8).
doi:10.14569/ijacsa.2017.080853.
[13] Terven, J. and Cordova-Esparza, D. (2023) A
comprehensive review of Yolo: From Yolov1 and beyond,
arXiv.org. Available at:
https://doi.org/10.48550/arXiv.2304.00501 (Accessed:
03 August 2023).
[14] Chi, J. et al. (2020) ‘A novel pulmonary nodule
detection model based on multi-step cascaded networks’,
Sensors, 20(15), p. 4301. doi:10.3390/s20154301.
[15] Potok, T.E. et al. (2018) ‘A study of complex deep
learning networks on high-performance, neuromorphic,
and quantum computers’, ACM Journal on Emerging
Technologies in Computing Systems, 14(2), pp. 1–21.
doi:10.1145/3178454.
[16] Jin, H. et al. (2023) ‘Machine learning techniques for
pulmonary nodule computer-aided diagnosis using CT
images: A systematic review’, Biomedical Signal
Processing and Control, 79, p. 104104.
doi:10.1016/j.bspc.2022.104104.
[17] Alzubaidi, L. et al. (2021) ‘Review of Deep Learning:
Concepts, CNN Architectures, challenges, applications,
Future Directions’, Journal of Big Data, 8(1).
doi:10.1186/s40537-021-00444-8.
[18] Ker, J. et al. (2018) ‘Deep learning applications in
medical image analysis’, IEEE Access, 6, pp. 9375–9389.
doi:10.1109/access.2017.2788044.

[19] Bhatt, D. et al. (2021) ‘CNN variants for Computer
Vision: History, architecture, application, challenges and
future scope’, Electronics, 10(20), p. 2470.
doi:10.3390/electronics10202470.
[20] Ünver, H.M. and Ayan, E. (2019) ‘Skin lesion
segmentation in dermoscopic images with combination of
Yolo and GrabCut algorithm’, Diagnostics, 9(3), p. 72.
doi:10.3390/diagnostics9030072.
[21] Tan, L. et al. (2021) ‘Comparison of RetinaNet, SSD,
and Yolo V3 for real-time pill identification’, BMC Medical
Informatics and Decision Making, 21(1).
doi:10.1186/s12911-021-01691-8.
[22] Ultralytics, Ultralytics/ultralytics: New - yolov8 in
PyTorch > ONNX > OpenVINO > CoreML >
TFLite, GitHub. Available at:
https://github.com/ultralytics/ultralytics (Accessed: 06
August 2023).

YOLOv8-Based Lung Nodule Detection: A Novel Hybrid Deep Learning Model Proposal

More Related Content

Similar to YOLOv8-Based Lung Nodule Detection: A Novel Hybrid Deep Learning Model Proposal (20)

More from IRJET Journal (20)

Recently uploaded (20)

YOLOv8-Based Lung Nodule Detection: A Novel Hybrid Deep Learning Model Proposal