A Comprehensive Review of Deep Learning

Methods for Lungs Cancer Detection Using 3D-

CNN
Ravinder Singh Shekhawat Navya Jain
Raj Kumar
Department of CSE Department of CSE
Department of CSE
Chandigarh University Chandigarh University
Chandigarh University
Mohali, India Mohali, India
Mohali, India
kumarraj.bgp2018@[Link]
Ishdeep Singla
Department of CSE
Chandigarh University
Mohali, India

Abstract— Lung cancer is a significant contributor to automatically extract key characteristics from training photos
global cancer-related mortality, responsible for approximately and conduct categorization. As a result, they act as integrated
one-quarter of all cancer-related deaths. It represents 13% of tools for lung cancer identification and hence eliminate the
newly diagnosed cancer cases and 19% of cancer-related need for complex image processing and pattern recognition
fatalities worldwide. Although complete prevention remains a pipelines in traditional techniques [5].
challenge, early diagnosis offers a crucial strategy for reducing
risks and improving patient survival rates. This review paper
introduces a novel approach utilizing deep learning,
specifically emphasizing the application of 3D Convolutional
Neural Networks (CNNs), to analyze patient-specific lung CT
scan data for potential lung cancer detection. The primary aim
is to extract valuable insights from a dataset comprising low-
dose CT scans, ultimately enabling the prediction of lung
cancer probability through the integration of advanced deep
learning techniques and 3D CNNs. The study envisions that
this research will empower the medical research community by
providing them with a comprehensive understanding of CNN
concepts, thereby catalysing progress in the broader field of
human healthcare. This review serves as a comprehensive
synthesis of existing knowledge, offering critical perspectives,
and laying the groundwork for future advancements in the
area of lung cancer detection using state-of-the-art deep Fig. 1. CT scan contains 5mm lung cancer nodule.
learning methodologies.
Convolutional Neural Networks (CNNs) are an important
Keywords—Nodule detection, Image segmentation, convolutional technique in deep learning for classifying things. Some
neural networks, Computer-aided diagnosis, segmentation. commonly used models, such as VGG (Visual Geometry
Group), are developed for flat pictures rather than three-
I. INTRODUCTION dimensional CT scans. Applying these techniques to 3D
photos require some changes. The third dimension is
Lung carcinoma is among the most prevalent regarded as an extra layer in the CNN, in which calculations
malignancies, contributing to upwards of 225,000 are conducted using a customized filter. This might make it
occurrences, 150,000 fatalities, and $12 billion in healthcare harder to see things in 3D. The three-dimensional CNN (3D-
expenditures on an annual basis within the United States [1]. CNN) was developed specifically for analyzing 3D pictures.
In the USA[18] and Europe[19], lung cancer is still the main It uses unique filters to interpret the three-dimensional
reason people die from cancer, making up about one-fifth structure.
(20.9%) of all cancer-related deaths. In accordance with the
World Health Organization (WHO), lung carcinoma
constitutes 2.2 million fresh instances each year, thereby II. RELATED WORK
representing 11.7% of all cancerous cases, in addition to 1.8 In Recent years we have witnessed a lot more growth in
million deaths, which is equivalent to roughly one-fifth of all the usage of deep artificial neural networks across various
cancer-related deaths, surpassing by a considerable margin domains of pattern recognition and machine learning, with
the combined figures of breast and colorectal carcinomas [2]. Convolutional Neural Networks (CNNs) occurring as a
Traditional CAD methods for lung cancer classification promising approach [6]. Here is one type named as CNNs,
involve various steps like image preprocessing, which demonstrated a outstanding performance in the 2012
segmentation, feature extraction, and classification [3]. The ImageNet Classification, surpassing established benchmarks
effectiveness of these systems is significantly impacted by in the computer vision community [7]. In the field of
the intermediary processes [4]. Deep learning approaches medical imaging, recent researches have found an deep
learning techniques, which is effectively generating Image Database Resource Initiative (LIDC/IDRI), which
encouraging outcomes and also providing some pretty good had around 1018 patients.
results.
In a study by R. Golan [12], they introduced a method to
Wu et al. were the first ones responsible for applying the fine-tune the settings of a special computer program
deep learning feature to enhance unequal registration of (Convolutional Neural Network or CNN) for finding lung
brain MR images by combining deep features [8]. Suk et al. nodules in CT scans. The system they tested correctly
suggested a new method that will use shared feature spotted nodules 78.9% of the time, but it also made 20
representation for neuro-imaging data, which will help in mistakes by flagging something as a nodule when it wasn't.
analyzing brain scans. It will be done using the Deep In another test, it identified nodules 71.2% of the time, with
Boltzmann machine for AD/MCI diagnosis[9]. Xu et al. 10 false alarms per scan. The study mainly looked at
have highlighted the importance of employing deep neural nodules that were agreed upon by all four radiologists.
networks to extract meaningful features from medical
images, which will be helpful in medical diagnosis and In recent research on standard computer vision datasets,
research[10]. Kumar et al. developed a computer program Convolutional Neural Networks (CNNs) have outperformed
that will be helpful in analyzing the lung CT scans and also Deep Belief Networks. Convolutional Neural Networks
identifying nodules. These use a special technique known as (CNNs) have become very popular in machine learning
an autoencoder, which will be helpful in learning the because they're really good at learning important things
characteristics of the nodules from a large dataset of labeled from the data they're given. This heightened interest stems
lung scans. This will also check whether the classification of from their potent representational capacity observed in
lung nodules is malicious or safe using the LIDC database. recent years.
[11]. Yaniv et al. created a system that uses a unique AI
named CNN to analyze the chest X-ray and find the
anomalies. In this system, they trained the AI using different III. CORRELATIONS AND INSIGHTS
images, which are even not medical X-rays. It will improve
the performance of the system, and they combined these A. Introduction
findings with another type of data analysis named PiCodes. Provide a brief overview of the reviewed works,
[12]. This new approaches shows the promising- advanced emphasizing their common focus on the application of deep
computer techniques to spot health issues in chest X-rays, learning techniques in medical imaging, particularly for tasks
even without using specialized medical data for training. related to lung nodule detection and classification.
The dataset included 93 images and achieved good results,
with a score of 0.93 for detecting Right Pleural Effusion, B. Key Themes and Overlapping Findings:
0.89 for identifying enlarged heart, and 0.79 for telling apart Feature Representation and Extraction: Discuss how
normal and abnormal chest X-rays. several works, including Suk et al. [8] and Xu et al. [10],
delve into methods for effective feature representation and
Tan et al. [10] conducted a study to develop a system extraction from medical imaging data. These approaches
that will be responsible for detecting lung nodules. They serve as foundational steps for subsequent tasks.
changed the method so that the false alarms could be
reduced. They used complicated computer techniques such Integration of Deep Learning Architectures: Highlight
as deep neural networks (DNN) and convolutional neural the efforts of Suna W. et al. [10] in exploring multiple deep
networks (CNN). CNN has a four-layer pattern for making learning algorithms, such as CNN, DBNs, and SDAE, and
simple information. They used simple, specific filters to their comparative analysis. This work showcases the
help this. They used data from around 85 patients obtained importance of considering various deep learning
from LIDC-IDRI. They also discovered a system that architectures for specific medical imaging tasks.
successfully identified 82% of the of the nodules of the
time. It also reduced the false alarms using DNN by Nodule Classification and Diagnosis: Examine the
approximately 33% [10]. contributions of Kumar et al. [11], J. Tan et al. [11], and R.
Golan [12], which primarily focus on lung nodule
Suna W. et al. [11] conducted a study in which they used
classification and diagnosis. Kumar et al. influence
There are three different computer methods for studying the
autoencoders for deep feature extraction, while J. Tan et al.
lung images. They are Convlutional Neural Networks
(CNN), Deep Belief Networks (DBNs), and Stacked and R. Golan apply CNN architectures with different
Denoising Autoencoders (SDAE). They compared these emphases.
three methods with a regular computer program that
Innovations in Neuro-Imaging: Acknowledge the work
searches for patterns in images. The CNN approach has
of Suk et al. [8] and Wu et al. [9], which pertain to neuro-
eight layers that search for different patterns. The regular
method used about 35 different features. After comparing, imaging data and brain function analysis. These studies
the CNN method was a bit better than the regular method, extend the application of deep learning beyond lung-related
with an accuracy of 0.7976, whereas the regular method has imaging tasks.
an accuracy of 0.7940. They used a large collection of lung
images known as the Lung Image Database Consortium and
 IV. ROLE OF 3D-CNN V. CHALLENGES AND FUTURE DIRECTIONS
Although there has been progress in applying 3D-CNNs
A. Architecture and Applications to identify lung cancer, there are still a number of issues that
A Convolutional Neural Network (CNN) is like a team of need to be resolved:
specialized filters that process images. It has different layers,
starting with filters that find patterns, then layers that connect 1) Data Availability: One major requirement to ensure
these patterns, and finally an output layer. The structure of a the accuracy of 3D-CNNs in detecting the presence of lung
CNN is like a recipe with layers stacked in a specific order, cancer is the utilization of large, properly structured and
alternating between finding patterns and simplifying labeled data sets. These datasets serve as the base for
information. To make a final decision, it uses a regular training the model to recognize important characterstics in
neural network with connected layers. lung scans, which allow more accurate detection of
abnormalities.
2) Computational Demands: Implementation of 3D-
CNN models requires substantial significant resources. This
include powerful hardware for both the training and
deployment phases. Just like a car needs a powerful engine
to run smoothly, this system relies on its computational
capacity for optimal performance.
Fig. 2. Architecture of a 3D convolutional neural network
3) Model Interpretability: It is important to know how
these 3D-CNNs make decisions, especially in healthcare
In Figure 2, you see a 3D shape on the left. After that, where things are very important. In order to use these
there are two layers that process information, followed by a technologies comfortably, medical professionals must
layer that connects everything, and at the end, there's the ensure that the technology provides correct diagnoses.
final result. In the layers that process information, each part Doctors are going to believe these technologies more since
that analyzes the data is shown as a shape. they have a better understanding of how they detect cancer.
3D Convolutional Neural Networks (3D-CNNs) are
being used in various task related to lungs cancer detection, Looking ahead, future research should focus on:
which include:
1) Developing Efficient 3D-CNN Architectures: In this
1) Nodule Detection: Nodule detection refers to area, the emphasis lies on optimizing 3D-CNN models to
identifying and selecting possible lung nodules on CT scans. require less computational power and storage resources.
Small, rounded, or oval-shaped nodules may be early signs Specialists are actively researching methods to minimize the
of lung cancer. Research by J. Tan et al. [11] and R. Golan computing power required by these systems while
[12] has shown that 3D-CNNs can accurately identify preserving their accuracy in detecting cancer. If these
nodules of various shapes and sizes. The regular systems require less computer power, they can be used on a
identification of these small abnormal growths is important wider range of computers. This means that more hospitals
for early detection, which in turn allows for more effective can use them, even if they lack the most powerful hardware.
treatment and improves the chances of positive result and 2) Advancing Interpretability: Researchers are working
prolonged survival. on ways that will provide clear insights into how 3D-CNN
2) Nodule Classification: After identifying nodules, the models arrive at decisions. The simplest way to do this is by
next step is to determine if they are malignant (cancerous) or using tools that show doctors the most important parts of the
benign. Sophisticated 3D convolutional neural network (3D- image that the system considered when making its decision.
CNN) models have proven to be an effective tool for This helps doctors understand how the system works better.
distinguishing between benign and malignant nodules, 3) Integrating 3D-CNNs into Clinical Workflows: For
thanks to intensive training on painstakingly labeled and any technology to be useful in healthcare, it has to work
classified [Link] nodular structures have proven to effortlessly with existing clinical processes. This implies
be extremely successfully classified using this sophisticated that implementing 3D-CNNs should aid radiologists and
machine learning technique. Precise categorization other medical professionals in performing their duties more
guarantees that patients have appropriate supplementary effectively, rather than complicating them. The aim is for
care and treatment regimens customized to the unique the technology to improve their abilities rather than restrict
features of the identified nodules, hence increasing the their workflow.
likelihood of a favorable consequence..
3) Cancer Staging: To determine if cancer has spread VI. CONCLUSION
beyond the lungs, additional tests are needed. The best In the end of my review paper on the use of 3D
treatment plan for a particular cancer type relies on its stage, convolutional neural networks (3D-CNNs) for lung cancer
indicating how extensively it has grown and spread detection, we identified that there are many transformative
throughout the body. This may involve undergoing MRIs or opportunities that they provide in terms of improved
PET scans to detect any metastasis. accuracy, automation, and non-invasive diagnostic
techniques. In this, it also includes the essential features
included in 3D-CNN studied by Han et al.'s [16] recent
work (2020) which revealed an incredible accuracy of
98.7% in identifying lung nodules using a 3D-CNN model. Neural Information Processing Systems 25 (NIPS 2012) (pp. 1097–
1105).
One of the most interesting parts was that it helped
[8] Suk, H., Lee, S., & Shen, D. (2014). Hierarchical feature
radiologists by supporting their own work in many ways. representation and multimodal fusion with deep learning for AD/MCI
This technique was really helpful in hospitals, and in many diagnosis. NeuroImage, 101, 569–582.
clinics where radiologists have less time, it also makes [9] Wu, G., Kim, M., Wang, Q., Gao, Y., Liao, S., & Shen, D. (2013).
fewer mistakes. In support of this, by reviewing the current Unsupervised deep feature learning for deformable registration of MR
CT images, 3DD-CNNs reduce the unwanted lung cancer brain images. Medical Image Computing and Computer-Assisted
Intervention, 16(Pt 2), 649–656.
identification. This approach significantly provides an
[10] Xu, Y., Mo, T., Feng, Q., Zhong, P., Lai, M., & Chang, E. I. (2014).
advantage over the regular approach, such as needle Deep learning of feature representation with multiple instance
biopsies, which cause discomfort and danger to patients. In learning for medical image analysis. In IEEE International
last, here the numbers show that 3D-CNNs are most Conference on Acoustics, Speech and Signal Processing, ICASSP
(pp. 1626–1630).
convenient and also promising and they have a lot of
[11] Kumar, D., Wong, A., & Clausi, D. A. (2015). Lung nodule
potential in future for finding lung cancer faster, which classification using deep features in CT images. In 2015 12th
means that patients can be diagnosed in a faster way as Conference on Computer and Robot Vision (pp. 133–138).
compared to previously. But, we should not stop here; it’s [12] Bar, Y., Diamant, I., Wolf, L., Lieberman, S., Konen, E., &
important to keep making this kind of CNN model and Greenspan, H. (2015). Chest pathology detection using deep learning
improve its accuracy as well, and doctors should also start with non-medical training. In Proceedings - International Symposium
on Biomedical Imaging (pp. 294–297).
using this new technology as it is easier to find using this
[13] Sun, W., Zheng, B., & Qian, W. (2016). Computer aided lung cancer
advanced technology. diagnosis with deep learning algorithms. In SPIE Medical Imaging,
9785, 97850Z–97850Z.
[14] Tan, J., Huo, Y., Liang, Z., & Li, L. (2017). A comparison study on
REFERENCES the effect of false positive reduction in deep learning based detection
for juxtapleural lung nodules: CNN vs DNN. In Proceedings of the
Symposium on Modeling and Simulation in Medicine, MSM ’17 (pp.
[1] Choi, W.-J., & Choi, T.-S. (2013). Automated pulmonary nodule
8:1–8:8).
detection system in computed tomography images: A hierarchical
block classification approach. Entropy, 15(2), 507–523. [15] Golan, R., Jacob, C., & Denzinger, J. (2016). Lung nodule detection
in CT images using deep convolutional neural networks. In 2016
[2] Amicizia, D., Arici, C., Biselli, R., Gasparini, R., & Panatto, D.
International Joint Conference on Neural Networks (IJCNN) (pp.
(2023). Systematic review of lung cancer screening: Advancements
243–250).
and strategies for implementation. Healthcare, 11(14), 2085.
[Link] [16] Han, H., Xu, J., Guo, Y., Xu, Z., & Chen, Y. (2020). A deep learning-
based framework for automatic lung cancer detection and
[3] Cheng, J. Z., Chou, Y. H., Huang, C. S., et al. (2010). Computer-
classification from CT images. IEEE Transactions on Medical
aided US diagnosis of breast lesions by using cell-based contour
Imaging, 39(7), 2175-2186.
grouping. Radiology, 255(3), 746-754.
[17] Alakwaa, W., Nassef, M., & Badr, A. (2017). Lung Cancer Detection
[4] Mahersia, H., Gabralla, L. A., & Zaroug, M. (2015). Lung cancer
and Classification with 3D Convolutional Neural Network (3D-
detection on CT scan images: A review on the analysis techniques.
CNN). International Journal of Advanced Computer Science and
International Journal of Advanced Research in Artificial Intelligence,
Applications/International Journal of Advanced Computer Science &
4(4), 38-45.
Applications, 8(8). [Link]
[5] Way, T. W., Sahiner, B., Chan, H. P., et al. (2009). Computer-aided
[18] Siegel, R.L., Miller, K.D., & Jemal, A. (2019). Cancer statistics,
diagnosis of pulmonary nodules on CT scans: Improvement of
2019. CA: A Cancer Journal for Clinicians, 69, 7–34. Retrieved from
classification performance with nodule surface features. Medical
Google Scholar [CrossRef] [PubMed]
Physics, 36(7), 3086-3098.
[19] Data Explorer ECIS. (2008). Retrieved August 5, 2008, from
[6] LeCun, Y., Kavukcuoglu, K., & Farabet, C. (2010). Convolutional
[Link]
networks and applications in vision. In Proceedings of the IEEE
1,2$3-All$6-0,14$5-2008,2008$7-8$CEstByCancer$X0_8-
International Symposium on Circuits and Systems (ISCAS) (pp. 253–
3$CEstRelativeCanc$X1_8-3$X1_9-AE28
256). IEEE.
[7] Alex, K., Sutskever, I., & Hinton, G. E. (2012). Imagenet
classification with deep convolutional neural networks. In F. Pereira,
C. J. C. Burges, L. Bottou, & K. Q. Weinberger (Eds.), Advances in