A Deep Learning Approach For Lung Cancer

Detection and Patient Support

Dr Fatma Mohamed Talaat
1
Department of Bio Artificial Intelligence , Faculty of Artificial Intelligence , Kafr El sheikh University , Egypt

Abstract: Lung cancer remains the leading cause of death worldwide.

Early and accurate detection is critical to improving patient outcomes.
Using complex algorithms and software, artificial intelligence (AI) is
able to mimic human cognition in analyzing, interpreting, and making
sense of complex data, and is currently being successfully applied in
various healthcare settings. By leveraging AI's ability to measure
information from images, and its superior ability to recognize complex
patterns in images compared to humans, AI has the potential to help
clinicians interpret LDCT images obtained in the lung cancer screening
setting. This book explores the potential of deep learning to
revolutionize lung cancer diagnosis. Intelligent Imaging: A Deep
Learning Approach to Lung Cancer Detection and Patient Support
details the development of a deep learning model to identify lung cancer
in chest scans. The book delves into the technical aspects of the model
while emphasizing its real-world applicability.

Introduction
Lung cancer is a malignant tumor that arises from lung cells, especially within the epithelial lining of the
bronchi, bronchioles, or alveoli. It is widespread and associated with high mortality rates on a global scale.
In its early stages, Lung cancer shows no symptoms or presents mild manifestations. As a result, it is usually
diagnosed at an advanced stage. Delay in identifying a medical condition affects the effectiveness of
treatment and reduces the likelihood of long-term survival. The two common types of Lung cancer are non-
small cell lung cancer (NSCLUNG CANCER) and small cell lung cancer (SCLUNG CANCER).
NSCLUNG CANCER is usually classified into two subtypes: lung squamous cell carcinoma (LUSC) and
lung adenocarcinoma (LUAD). The precise classification of Lung cancer, including LUSC, LUAD, and
SCLUNG CANCER, has an important role in determining the prognosis of Lung cancer compared with
benign and malignant classifications. Accurate classification of LUNG CANCER at the initial diagnosis
stage significantly improves the treatment efficacy and thus increases the survival rate of patients. PET and
computed tomography (CT) are widely used as non-invasive diagnostic imaging modalities in clinical
practice, where they serve as valuable tools to evaluate the specific diagnosis of Lung cancer.
Early detection and treatment of Lung cancer through effective screening methods is vital in enhancing
patient outcomes. Based on results from the National Lung Screening Trial, low-dose helical CT screening
is more effective in reducing mortality in high-risk populations. However, the Lung cancer screening
process is prone to giving false positive (FP) results, which leads to increased costs due to unwarranted
medical interventions and may lead to psychological distress in individuals. Computer-aided diagnosis has
notable advantages in Lung cancer detection, including enhanced scope in early cancer screening and
reduced incidence of FP findings throughout the diagnostic process. In the field of Lung cancer detection,
there have been notable developments in the form of new methods and technologies to enhance early
diagnosis and treatment effectiveness. Liquid biopsies are used to test blood samples for cancer. These
diagnostic tests can identify genetic abnormalities and alterations associated with Lung cancer. Thus, these
tests provide a non-invasive way to diagnose the disease and monitor the effectiveness of treatment. Low
dose computed tomography (LDCT) scanning has emerged as a widely applied method for timely
identification of Lung cancer. LDCT scans use lower radiation levels than traditional CT scans while
providing high-resolution images of the lung area. New bronchoscopy methodologies, including
electromagnetic navigation bronchoscopy and robot-assisted bronchoscopy, facilitate minimally invasive
lung lesion biopsies. These tools promote rapid and accurate identification of medical conditions. The
integration of genomics, proteomics and metabolomics has allowed the development of diverse strategies
for Lung cancer identification. These methodologies use several molecular markers to enhance diagnostic
accuracy and discover potential targets for therapeutic intervention.
Deep learning-based Lung cancer screening techniques can reduce mortality by detecting the disease at the
initial stage. It can help reduce false negatives (FN) by detecting subtle or early indicators of Lung cancer
that humans may leave unnoticed. Imaging modalities including computed tomography, magnetic resonance
imaging (MRI), and computed tomography (PET) can be combined using DL algorithms to understand the
disease and aid in diagnosis and treatment planning. The staging of cancer is directly related to the extent
of the disease. A combination of imaging methods and biopsies of suspicious tissue identifies types of
cancer. Cancer staging helps caregivers choose chemotherapy, immunotherapy, radiation, and surgical
strategies. In particular, a higher cancer stage increases the mortality rate. The effectiveness of medical
treatments depends on the stage of the cancer. Providing more accurate and trustworthy diagnoses can help
reduce misdiagnosis and unwarranted care. The Lung cancer scan process produces large amounts of
medical imaging data. DL models effectively analyze CT scans, chest radiography, and other imaging
modalities due to their ability to analyze big data. Using DL models, doctors can gain insight into a patient's
condition by combining data from multiple imaging modalities and other clinical data sources.
Due to privacy concerns and the cost of data acquisition, medical image datasets are typically small. Pre-
trained models can apply broader image dataset expertise to medical images, enabling models to be trained
with minimal medical data. These models can extract hierarchical information from images, including fine
details and important patterns. Feature extraction supports medical image classification model

Figure 1:CT scan slice containing a small (5mm) early-stage lung cancer nodule [3].
1. BACKGROUND

The process of diagnosing some diseases can be aided using automation methods through computer-aided
diagnosis (CAD). This method uses software to segment, predict, localize and classify symptoms, which
will be used to infer the presence and severity of diseases. In this work, the focus is on providing a review
of CAD methods used to identify cancerous nodules in lung CT scans. In general, CT allows doctors to
recognize the presence of lung cancer nodules, especially when the nodules are large, which corresponds to
the late stage of the disease. However, it is important to recognize nodules at an early stage, which are
usually small in size before the patient develops a golf ball-sized lung tumor. Several studies have
investigated the use of deep learning algorithms for CT-based lung cancer screening and diagnosis. In
general, there are unique image attenuation patterns in CT images for healthy and unhealthy scans. To
distinguish the lungs from the surrounding tissues, straightforward techniques such as numerical
approaches, gray-level thresholding, and shape-based approaches can be used to perform simple lung
segmentation.

Several studies have demonstrated the efficacy of deep learning in lung cancer detection:

- Biradar et al. proposed a 2D CNN algorithm achieving 88.76% accuracy in classifying lung nodules
from CT images.

- Zhang et al. combined CNN with morphological features for nodule detection, achieving high accuracy in
recognizing pulmonary nodules.

- Ramkumar et al. utilized AI and deep learning techniques for tumor evaluation, emphasizing the role of
supervised learning in improving detection performance.

2. Classification of CT scans
Data contains 3 chest cancer types which are Adenocarcinoma, Large cell carcinoma, Squamous cell
carcinoma , and 1 folder for the normal cell Data folder is the main folder that contains all the step folders
inside Data folder are test , train, valid.
The precise location and size of a lung tumor can depend on factors such as the specific subtype of cancer,
the stage of the disease, and the patient's unique anatomy.
Our primary dataset is the patient lung CT scan dataset from Kaggle. The dataset contains labeled data for
13362 patients, which we divide into training set of size 11494, validation set of size 1553, and test set of
size 315. For each patient the data consists of CT scan data and a label (0 for no cancer, 1 for cancer). Note
that the Kaggle dataset does not have labeled nodules. For each patient, the CT scan data consists of a
variable number of images (typically around 100-400, each image is an axial slice) of 512×512 pixels. The
slices are provided in DICOM format. Around 70% of the provided labels in the Kaggle dataset are 0, so
we use a weighted loss function in our malignancy classifier to address this imbalance.
 (a) (b) (c)
Figure 1: CT scans for lung cancer.

2.1 preprocessing
we made a preprocessing on our data to make sure that the model will understand the data.
Resize Images: MobileNet, like most CNNs, expects input images to have a fixed size. The size can vary
depending on the specific MobileNet variant you're using (e.g., MobileNetV1, MobileNetV2). For example,
MobileNetV1 commonly uses 224x224 pixel images, while MobileNetV2 can handle various input sizes.
Mean Subtraction: Similar to VGG16, you may subtract the mean pixel value across all images in the
dataset. However, MobileNet models might have been trained with different datasets than VGG16, so it's
essential to use the appropriate mean pixel values for MobileNet.
Normalization: Normalize the pixel values to a specific range. MobileNet typically expects pixel values in
the range [-1, 1] or [0, 1]. You might need to scale the pixel values accordingly.
Channel Ordering: Ensure that the input image is in the correct channel ordering expected by the model.
MobileNet typically expects images to be in the 'RGB' format.
Batching: Prepare the input data in batches for efficient processing. MobileNet, like VGG16, benefits from
processing images in batches for parallel processing.

3.METHODOLGY
We used MobileNet model MobileNet is a computer vision model open-sourced by Google and designed for
training classifiers. It uses depth wise convolutions to significantly reduce the number of parameters compared to
other networks, resulting in a lightweight deep neural network. MobileNet is TensorFlow’s first mobile computer
vision model.
Figure 2 : Mobilenet Architecture

Our approach involves a CNN-based model, referred to as CNN-ALCD (Automatic Lung Cancer
Detection), which leverages supervised learning to detect lung cancer from CT scan images. The
methodology includes several key steps:

1. Data Collection and Preprocessing:

- Data Collection: The dataset comprises low-dose computed tomography (LDCT) images sourced from
multiple medical institutions.
- Preprocessing: Enhancing image quality through normalization, augmentation, and segmentation to
highlight relevant features.

2. Model Architecture:
- Convolutional Layers: The CNN-ALCD model consists of multiple convolutional layers designed to
extract high-level features from the input images.
- Pooling Layers: Max-pooling layers are used to reduce the dimensionality of the feature maps, retaining
only the most relevant information.
- Fully Connected Layers: These layers are used to classify the extracted features into cancerous and non-
cancerous categories.
- Activation Functions: ReLU activation functions are used to introduce non-linearity into the model, while
the softmax function is used in the output layer for classification.

3. Training and Validation:

- Training: The model is trained using a large dataset with labeled CT images, optimizing the parameters
using the Adam optimizer and cross-entropy loss function.
- Validation: A portion of the dataset is set aside for validation to ensure the model generalizes well to
unseen data. Performance metrics such as accuracy, sensitivity, and specificity are used to evaluate the
model.

4. Implementation:
- The model is implemented using Python and libraries such as TensorFlow and Keras. It is deployed on
cloud platforms to ensure scalability and accessibility.
2.2 LUNG CANCER DETECTION
It is used to discriminate between multiple image sets. Image classification is the process of classifying
and labeling pixel sets within an image depending on certain rules. The classification law can be formulated
using one or more spectral or textural features and properties. The two major methods of classification are
supervised and unsupervised. Supervised learning is a machine learning approach that depends on labeled
datasets. These datasets are designed to train algorithms to accurately classify data or predict outcomes. The
model can evaluate its accuracy and learn over time by using input and output labeled data. Unsupervised
learning uses the algorithms of the machine learning for analysis and clustering of unlabeled datasets. These
algorithms discover he unseen patterns in data without the need for human intervention [14].

Figure 4: Block diagram of classification of CT Scan images.

The block diagram of the proposed method of classifying CT images into three classes i.e.
viral, COVID, and normal. Firstly, lung images are entered into the AlexNet architecture for
feature extraction, and then by using SVM for more accurate classification, being viral, COVID,
or normal of each image is determined precisely and automatically.
 Figure 5: Block diagram of lung cancer detection.

Results
The proposed CNN-ALCD model outperformed existing neural network-based methods, achieving
an accuracy rate of 94.11%. The confusion matrix analysis revealed a high true positive rate,
indicating the model's potential in clinical settings.
 3 .Conclusion

Lung cancer remains a leading cause of mortality worldwide, primarily due to its late-stage diagnosis. Early
and precise detection significantly improves treatment efficacy and patient survival rates. This paper
explores the potential of deep learning, particularly convolutional neural networks (CNNs), in
revolutionizing lung cancer diagnosis. By leveraging advanced imaging techniques and AI algorithms, it is
possible to identify lung cancer at its initial stages, thus reducing false negatives and enhancing diagnostic
accuracy.

Techniques such as low-dose computed tomography (LDCT) and computer-aided diagnosis (CAD) are
crucial in early cancer screening. LDCT provides high-resolution images with lower radiation, while CAD
enhances early diagnosis accuracy, reducing false positives. Additionally, integrating genomics, proteomics,
and metabolomics with deep learning models allows for a comprehensive analysis, aiding in accurate
classification and staging of lung cancer.

The deployment of pre-trained models on limited medical datasets, combined with advanced preprocessing
techniques, facilitates the extraction of detailed information from medical images. This approach supports
the development of robust classifiers that can handle the complexities of lung cancer diagnosis.

In summary, deep learning offers a promising avenue for improving lung cancer detection and patient
outcomes by providing early, accurate, and reliable diagnostic tools.

References

1. Biradar, V. G., Pareek, P. K., & Nagarathna, P. (2022). Lung Cancer Detection and Classification
using 2D Convolutional Neural Network. IEEE MysoreCon.

2. Zhang, Y., Dai, B., Dong, M., Chen, H., & Zhou, M. (2022). A Lung Cancer Detection and Recognition
Method Combining Convolutional Neural Network and Morphological Features. IEEE CCET.

3. Ramkumar, M., Babu, C. G., Wahhab, A. R. A., Abinaya, K., Balaji, B. A., & Chakravarthy, N. A. (2022).
Detection and Diagnosis of Lung Cancer using Machine Learning Convolutional Neural Network
Technique. IEEE STCR.