![](/jspui/image/dspace_header.jpg)
Please use this identifier to cite or link to this item:
https://idr.nitk.ac.in/jspui/handle/123456789/17721
Title: | Deep Learning-Based Decision Support System For Lung Cancer Detection |
Authors: | Jagdish, Dodia Shubham |
Supervisors: | B., Annappa |
Keywords: | Lung cancer;Medical imaging;Nodule segmentation;Nodule detection |
Issue Date: | 2023 |
Publisher: | National Institute Of Technology Karnataka Surathkal |
Abstract: | Cancer is a major cause of significant fatal rates and morbidity worldwide. According to the latest World Health Organization (WHO) estimates issued in 2020, cancer disease has the greatest mortality rate, accounting for around 10 million deaths. Over a lifetime, 1 in 18 men and 1 in 46 women are known to develop lung cancer. Accurate identification of lung cancer has been a challenging task for decades. Even though there are techniques to identify lung cancer nodules, it takes enormous efforts from expert radiologists. Therefore, it is very crucial to automate the process of identifying nodules from Computed Tomography (CT) scans. This thesis discusses the methods proposed to perform lung cancer detection, segmentation, and classification using novel deep learning algorithms. First, the task of detecting lung nodules from CT scan images is performed. The lung nodules are irregular tissue formations that can be as small as 3 mm in diameter. The detection of these lung nodules is a tedious and time-consuming task, as careful examination needs to be carried out by radiologists. The annotations that the radiologists provide must be precise and accurate as well. This can lead to human error. Combining this with computer-aided algorithmic solutions may resolve this issue. However, deploying this real-time environment is another challenge as it needs to be interfaced with these solutions as per doctor’s requirements. In this thesis, different deep-learning solutions are used to develop lung cancer nodule detection from CT scan images. The potential nodule candidates are identified by the proposed detection methods. Second, the task of segmenting the nodule regions from the detected lung nodules is performed. Once the potential nodule candidates are detected, the accurate nodules are to be segmented. One of the main challenges that occur in segmentation of lung nodules is that non-nodules that appear like nodules can be segmented. Therefore, segmentation of lung nodules is essential to avoid misdiagnosis. In this thesis, Artificial Intelligence(AI)-based methods are proposed to perform accurate lung nodule segmentation tasks from the input CT scans. Third, the task of classifying a segmented nodule as cancerous or non-cancerous is performed. The tumor/nodule found in the lung/thoracic region can be malignant or benign. The spread rate and re-occurrence of a malignant nodule in the human body are very rapid. Therefore, it is crucial to identify the type of nodule at the earliest. In this thesis, various deep-learning solutions are designed and developed to perform this task. All the proposed methods in this thesis are evaluated on the publicly available benchmark LUNA16 dataset; their respective results are presented in subsequent chapters and verified by an expert pulmonologist. The proposed models resulted in superior performance in comparison with state-of-the-art techniques. When compared, state-of-the-art techniques had accuracies of 96.9%, 94.97%, and 96.9% for the detection, segmentation, and classification task, respectively. However, the proposed models yielded an accuracy of 98.21% for the detection task, a dice similarity coefficient of 98.0% for the segmentation task, and an accuracy of 98.7% for the classification task. This clearly shows an improvement of 1.31%, 3.03%, and 1.8% for the detection, segmentation, and classification tasks respectively. |
URI: | http://idr.nitk.ac.in/jspui/handle/123456789/17721 |
Appears in Collections: | 1. Ph.D Theses |
Files in This Item:
There are no files associated with this item.
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.