Publication: Covid-19 Tanısında Konvolüsyonel Sinir Ağları Mimarilerinin Karşılaştırılması
Abstract
Amaç: COVID-19'un bulaş hızının önüne geçebilmek için erken tanının yüksek doğrulukla yapılması esastır. Akciğerde görülen buzlu cam görünümü ve doku değişiklikleri revers-transkriptaz polimeraz zincir reaksiyonu (RT-PCR) sonuçları negatif olan hastalarda da karşılaşılan klinik sempomlar arasında olup, göğüs bilgisayarlı tomografi (BT) görüntüleri ile rahatlıkla farkedilebilmektedir. Bu çalışmanın amacı, BT görüntüleri üzerinden, COVID-19'un doğru ve hızlı tanısına olanak sağlayan, konvolüsyonel sinir ağları (KSA) mimarileri kullanılarak hastalığın belirlenmesi ve mimarilerin çeşitli performans ölçüleriyle sınıflandırma başarılarının karşılaştırılarak değerlendirilmesidir. Materyal ve Metot: Çalışma verisinde, çalışma kriterlerine uygun olarak seçilen, RT-PCR test sonucu olan ve BT'si çekilen 363 pozitif vakadan 8405 kesit, 134 negatif vakadan 7189 kesit olmak üzere, toplamda 15594 akciğer BT kesiti elde edildi. Veriler, KSA mimarilerinden DenseNet169, MobileNetV2, ResNet50, InceptionResNetV2, InceptionV3, VGG16, VGG19, Xception, DenseNet121, DenseNet201 ve ResNet101 kullanılarak analiz edildi. Veri seti eğitim-test olarak %80-%20 oranında bölündü. Performans değerlendirmesi için doğruluk, duyarlık, seçicilik, kesinlik, F1 skoru, ROC eğrisi ve AUC performans ölçüleri kullanıldı. Veri analizleri Python programlama dili ile gerçekleştirildi. Python programlama dili TensorFlow ve Keras kütüphaneleri ile birlikte kullanıldı. Bulgular: Çalışma verisi ile en yüksek performansı gösteren mimari %99,13 doğruluk, %99,86 duyarlık, %98,52 seçicilik, %98,28 kesinlik, %99,07 F1 skoru ve %99,19 AUC değerleri ile DenseNet201 olurken, en düşük performansa sahip mimari %92,47 doğruluk, %99,79 duyarlık, %86,25 seçicilik, %86,03 kesinlik, %92,40 F1 skoru ve %93,02 AUC değerleri ile VGG19 oldu. Sonuç: Sonuç olarak bu çalışma ile, kullanılan mimarilerin genelleştirme yeteneklerinin iyi olduğu ve görüntü verileri üzerinden yapılan sınıflandırmalarda klinik problemlere cevap verebildikleri saptandı. Bu çalışmada COVID-19'un BT görüntüleri kullanılarak tanısında, KSA yönteminin, karar alıcılara yardımcı olacak bir karar mekanizması olduğu ve bu sayede hastalığa en kısa sürede tanı konulabileceği, tedavi etkinliğinin artırılabileceği, hastalığın yarattığı olumsuzlukların önüne geçilebileceği ortaya konulmuştur. Anahtar Sözcükler: COVID-19, Derin öğrenme, Konvolüsyonel sinir ağları, Görüntü sınıflandırma, Bilgisayarlı tomografi
Aim: In order to prevent the transmission rate of COVID-19, early diagnosis with high accuracy is essential. The ground glass opacity and tissue changes in the lung are among the clinical symptoms encountered in patients with negative reverse transcription-polymerase chain reaction (RT-PCR) results and can be easily recognised by chest computed tomography (CT) images. The aim of this study is to identify the disease using convolutional neural network (CNN) architectures that allow accurate and rapid diagnosis of COVID-19 on CT images and to evaluate the classification success of the architectures by comparing their classification success with various performance metrics. Materials and Methods: In the study data, a total of 15594 lung CT slices were obtained, 8405 slices from 363 positive cases and 7189 slices from 134 negative cases, selected in accordance with the study criteria, with RT-PCR test results and CT scans. The data were analysed using CNN architectures DenseNet169, MobileNetV2, ResNet50, InceptionResNetV2, InceptionV3, VGG16, VGG19, Xception, DenseNet121, DenseNet201 and ResNet101. The dataset was split 80%-20% as train-test. Accuracy, sensitivity, specificity, precision, F1 score, ROC curve and AUC performance metrics were used for performance evaluation. Data analyses were performed with Python programming language. Python programming language was used together with TensorFlow and Keras libraries. Results: The architecture with the highest performance with the study data was DenseNet201 with accuracy of 99.13%, sensitivity of 99.86%, specificity of 98.52%, precision of 98.28%, F1 score of 99.07%, and AUC of 99.19%, while the architecture with the lowest performance was VGG19 with accuracy of 92.47%, sensitivity of 99.79%, specificity of 86.25%, precision of 86.03%, F1 score of 92.40%, and AUC of 93.02%. Conclusions: As a result, it was determined that the architectures used in this study have good generalization capabilities and can respond to clinical problems in classifications made on image data. In this study, it has been demonstrated that the CNN method is a decision mechanism that will help decision makers in the diagnosis of COVID-19 using CT images, and in this way, the disease can be diagnosed as soon as possible, treatment effectiveness can be increased, and the negativities caused by the disease can be prevented. Keywords: COVID-19, Deep learning, Convolutional neural networks, Image classification, Computed tomography
Aim: In order to prevent the transmission rate of COVID-19, early diagnosis with high accuracy is essential. The ground glass opacity and tissue changes in the lung are among the clinical symptoms encountered in patients with negative reverse transcription-polymerase chain reaction (RT-PCR) results and can be easily recognised by chest computed tomography (CT) images. The aim of this study is to identify the disease using convolutional neural network (CNN) architectures that allow accurate and rapid diagnosis of COVID-19 on CT images and to evaluate the classification success of the architectures by comparing their classification success with various performance metrics. Materials and Methods: In the study data, a total of 15594 lung CT slices were obtained, 8405 slices from 363 positive cases and 7189 slices from 134 negative cases, selected in accordance with the study criteria, with RT-PCR test results and CT scans. The data were analysed using CNN architectures DenseNet169, MobileNetV2, ResNet50, InceptionResNetV2, InceptionV3, VGG16, VGG19, Xception, DenseNet121, DenseNet201 and ResNet101. The dataset was split 80%-20% as train-test. Accuracy, sensitivity, specificity, precision, F1 score, ROC curve and AUC performance metrics were used for performance evaluation. Data analyses were performed with Python programming language. Python programming language was used together with TensorFlow and Keras libraries. Results: The architecture with the highest performance with the study data was DenseNet201 with accuracy of 99.13%, sensitivity of 99.86%, specificity of 98.52%, precision of 98.28%, F1 score of 99.07%, and AUC of 99.19%, while the architecture with the lowest performance was VGG19 with accuracy of 92.47%, sensitivity of 99.79%, specificity of 86.25%, precision of 86.03%, F1 score of 92.40%, and AUC of 93.02%. Conclusions: As a result, it was determined that the architectures used in this study have good generalization capabilities and can respond to clinical problems in classifications made on image data. In this study, it has been demonstrated that the CNN method is a decision mechanism that will help decision makers in the diagnosis of COVID-19 using CT images, and in this way, the disease can be diagnosed as soon as possible, treatment effectiveness can be increased, and the negativities caused by the disease can be prevented. Keywords: COVID-19, Deep learning, Convolutional neural networks, Image classification, Computed tomography
Description
Keywords
Citation
WoS Q
Scopus Q
Source
Volume
Issue
Start Page
End Page
85
