الفهرس 
Chapter One: IntroductionOnly 14 pages are availabe for public view
 |  | from | 137 |  |  |
| | | from | 137 | | |
Abstract
Overview
This chapter summarizes the results and findings of this dissertation and draw some conclusions. In addition, it also presents the potential avenues of research and future studies.
5.2 Summary
Early diagnosis of lung nodules in CT scans is of great importance for successful treatment and decreasing the mortality rates in affected patients. Due to improvements in screening technologies, and an increased demand for their use, radiologists are required to analyze an ever-increasing amount of image data, which can affect the quality of their diagnoses. Many CAD systems are designed to assist radiologists in this endeavour.
CNNs have proven to be well suited to classify lung nodules due to its capacity to learn complicated features from medical images. However, further optimization is required to make them as precise and broadly applicable as feasible. The concepts of natural selection and evolutionary theory such as GA serve as inspiration.
the aim of this dissertation is to introduce a novel automated algorithms for 3D CNNs architecture design based on GA or DNA computing to classify whether pulmonary nodules are malignant or benign from 3D CT scans, without needing expert knowledge on feature extraction or CNN design. Furthermore, examine the potential of the proposed approaches to predict the level of cancerous of individuals with lung cancer based on several symptoms’ features.
Classifying the candidates from lung CT images is very important for radiologists since analysing these scans is a heavy burden for them that for a single scan, there are many slices to be reviewed where it takes time and is prone to error to review these slices. Therefore, these scans are occasionally evaluated by many radiologists to obtain a strong result. That why it is extremely essential cure to help the radiologists by lowering the quantity of applicants to review. Also, one of the primary reasons for the low survival rates following a lung cancer diagnosis is the absence of efficient early detection diagnostic techniques. Therefore, there is significant value in conducting research in this domain to explore novel approaches that could enhance the diagnosis condition using the patients’ clinical history and medical symptoms such as cough, shortness of breath, fatigue, weight loss and chest pain.
For image classification we utilize the dataset of LUNA 2016 for evaluation with both algorithms and for level prediction we used a public dataset for symptoms’ features. The proposed approaches mainly consist of two steps: pre-processing and building classification model using GA or DNA computing. In image classification the pre-processing step, is a series of steps involving the collection, normalizing, and cropping 3D data resulting a total of 874 lung nodules. By utilizing both approaches in the second step, we were able to automatically design 3D CNN architectures for pulmonary nodule image classification and compare the effects of different patch sizes when training the algorithms to increase overall performance. On other hand we also, exploit the proposed algorithms to develop models for predicting the level of cancerous of individuals with lung cancer manifestation based on several features–symptoms using 1D CNN after preparing the input data by cleaning, transforming, and eliminating ineffective features.
5.3 Conclusions
In term of the 3D lung nodule images classification, the goal was achieved successfully by developing cutting-edge approaches which exploits the GA and the DNA computing search to encode and design 3D CNNs with random depths as both approaches:
• Accept a 3D image as input.
• Evaluate the algorithm on the LUNA16 dataset using 4-fold cross-validation.
• Enable us to automatically extract the best features for lung nodule classification.
• Provide 3D CNN architectures without the need to manually investigate.
In conclusion both approaches provide an end-to-end methodology from LUNA16 dataset pre-processing to 3D CNN architecture optimization to final classification by using the GA approach or the DNA computing approach to optimize CNN architecture hyperparameters. Both approaches achieved a state-of-the-art classification performance, with accuracy up to 97.7% on the test set and compare favorably to other methods with a compact model size of just 1.18 million parameters.
The results obtained from the proposed models demonstrate their effectiveness in accurately classifying pulmonary nodules. Achieving high accuracy, sensitivity, specificity, and precision underscores the potential of these models for enhancing early detection and diagnosis in lung cancer cases. The robust performance, as indicated by the ROC curve’s high AUC value, further solidifies the models’ discriminative capabilities. Overall, the proposed algorithm exhibits promise as a valuable tool in medical image analysis for pulmonary nodule identification.
In general, the main advantages of the proposed algorithms are as follows.
• Dealing with 3D images rather than 2D images that provides additional benefits such as:
o Better feature capture: in a 3D image, the algorithm has access to data along three axes (X, Y, and Z), which may allow for more robust feature extraction. In contrast, 2D images provide data along only two axes, potentially missing important characteristics.
o Reduced data loss: converting 3D structures to 2D images often results in a loss of data. With 3D imaging, the integrity of the original data is maintained.
o More accurate representation of real-world objects: 3D images capture depth, volume, and spatial relationships better than a flat 2D image. This allows for a more detailed and realistic representation.
• Rather than having to manually investigate and extract features, both algorithms enable us to automatically extract the best features.
• The automated output of CNN architecture that can be related to the input data and does not require the presence of experts.
The disadvantages of the proposed algorithms are as follows.
• Computational complexity: the proposed algorithms require significant computational resources to run, particularly when dealing with large datasets. This can make them slow and computationally expensive.
• Quality of the final solution is not guaranteed: The quality of the final solution obtained by the proposed algorithms is not guaranteed to be optimal or the same every time. The algorithm may converge to a solution that is not optimal, depending on the implementation.
In term of cancer level prediction both approaches showed robust performance with the absence of misclassifications which demonstrate the robustness performance in distinguishing between different levels. The attained precision, recall, and F1 score values of 1 underscore the models’ outstanding ability to accurately classify instances across different levels. These results instill confidence in the models’ reliability and effectiveness in the given classification task.