Razi Journal of Medical Sciences

Mendeley  

Zotero  

Ghahvechi Khaligh, H, Pourasad Y, Moghadas Gholian S. Classification of Lung nodules using textural and geometric features. RJMS 2020; 27 (4) :106-121
URL: http://rjms.iums.ac.ir/article-1-6145-en.html

Background: Background: Since the diagnosis of cancerous and malignant lung glands using imaging techniques such as CT-Scan without the need for sampling reduces the risk of spreading cancerous nodules, the development of a computerized diagnostic system for processing images and pulmonary glands and then class Their classification into two categories, benign and malignant, plays an important role in the early diagnosis of lung cancer and the survival of patients. Access to a database with a uniform statistical population of malignant and benign glands is one of the most fundamental steps in the implementation and evaluation of computerized diagnostic systems for cancer patients. In the present study, the image database consortium image collection of lung images has been used. This database includes images of CT scans of lung cancer, along with the diagnostic opinion of a specialist doctor and the identified areas of the glands. This database has been compiled by the National Cancer Organization of America, the National Health Organization and the Food and Drug Administration made available to the public. In this database, the CT images of each patient are stored in DICOM format, which is the standard for storing and suitable for processing medical images. The average incision of each scan for each patient is 254 incisions, the distance between each incision is 9 to 9 mm. The aim of this study is to achieve higher classification accuracy and therefore higher diagnostic accuracy of malignant and benign glands.
Methods: In this study, the algorithms that have been used to classify the pulmonary glands are introduced and finally the proposed algorithm is presented. In the proposed algorithm, the CT scan images of the lungs are pre-processed and then extracted from the nodule area by the active Chen-Wess contour. From the fragmented area, the histogram, texture and geometric features are extracted. These features then classify the pulmonary nodules into two categories, benign and malignant, using two classes, SVM and KNN. After extracting the features from the fragmented areas and normalizing them, with a large amount of data (feature) we are faced with using this data to make the final decision and classification about whether the glands are benign or malignant due to the large number. It is necessary to choose the best and most valuable features for the correct classification. There are different ways to select the feature, but due to the time consuming nature of this process, in the present study, this step is eliminated and first all the features are classified. Then, by trying and making a mistake, the best features are selected for each class. Therefore, by doing this, the feature and classification are selected at the same time and the computational load and processing time are reduced. In this research, the extracted features is given to the two well-known classifiers of the support vector machine and the nearest neighbor parameter. This database has three Excel files, the first / adjacent file contains 6 information such as the number of nodules in each patient along with the size of each nodule, the main source of the nodule and the final diagnosis of the radiologist and the treatment for each patient. The second file contains general information such as the date of each scan, the name of the company that made the CT device, the device model, the software version, and the image ID. In the third Excel file, the number of nodules larger than 9 mm and smaller than 9 mm for each patient is given.
Results: Each scan was examined separately by four radiologists, and scans identified by all four radiologists were added to the database. Experts have divided each nodule into one of four unknown categories: benign, benign, primary malignant, and metastatic malignancy. In this study, for each patient, the incision in which the nodule appears is selected. Also, since the nodules are classified into 4 categories, in this study, we have classified the unknown and benign category as benign and the primary malignant and malignant metastatic categories as malignant. In this study, lung images of 65 cancer patients from the mentioned database were classified, of which 49 patients had malignant nodules and 25 patients had benign nodules. In the proposed algorithm, a semi-automatic method is used to segment the pulmonary nodule area. Using an automatic classification algorithm that does not require the selection of two border points of the gland, it can increase the rate of fragmentation. Of course, the advantage of the semi-automatic segmentation method is its high classification accuracy, which is much lower in automatic segmentation. Therefore, proposing and implementing an automated segmentation algorithm that is both highly accurate is at the top of the project's future plans. In this study, the lung glands were classified into two categories: benign and malignant. The results of the proposed actions are examined. The pre-processed images are then aligned by the Chen-and-Three algorithm and the area is extracted and subjected to feature extraction algorithms, and 25 different tissue and geometric features are extracted for each gland from these areas. Finally, by extracted data, the SVM and KNN classifications classify the glands. Criteria for accuracy, sensitivity and specificity in the top class are obtained by 90.8%, 100% and 89%.
Conclusion: In addition to high accuracy in diagnosis, this method is also a low cost and low risk method. By comparing the results of the proposed method with the previous methods, the proposed method received the most sensitivity, and in many studies, the highest classification accuracy. Given that the criterion of sensitivity means the ability of classification in the correct diagnosis of the disease in a person and the criterion of specificity means the ability of classification in the correct diagnosis of the absence of the disease in the person, so it can be concluded that the criterion Sensitivity is very important in research related to glandular diagnosis in medical imaging. This is because correctly diagnosing the presence of a disease or cancer is much more important and vital than diagnosing its absence in a candidate. Therefore, the proposed method is recommended as an efficient and suitable method for classification of the pulmonary glands due to its very high sensitivity and also having the desired values of two criteria of accuracy and specificity and low number of features used for classification.