Deep Transfer Learning Based Approaches for Classification of Chest CT Scans: Normal, COVID-19, and Pneumonia

Since the start of 2020, the global community has been shocked by the coronavirus disease(COVID-19). The most popular method for diagnosing COVID-19 is RTPCR testing, which is very time-consuming and requires isolating environment. On the other hand, radiography images has been proven to be beneficiary in automated and early diagnosis of COVID-19, which can reduce the fatal transmission as well as mortality rate. In this research work, transfer learning was employed where InceptionV3, VGG19 and ResNet50 were used as pre-trained models. These three convolutional neural network(CNN) architectures were initially trained on ImageNet dataset and then Large COVID-19 CT scan dataset for training, testing and validation. The dataset went through several pre-processing and augmentation steps like - image resizing and rescaling followed by rotation, brightness adjustment, horizontal and vertical flipping and normalization. Subsequently, feature extraction was performed using the three pre-trained CNN architectures and additional layers were introduced including flattening, dense, batch normalization, and dropout for further refinement. Lastly, the images were classified into three classes - Normal, Covid-19 and CAP(Community Acquired Pneumonia) in the last dense(fully connected) layer with a softmax activation function. Hyperparameter tuning was undertaken to maximize the performance of the models. The comparative analysis of the performance of InceptionV3, VGG19 and ResNet50 was presented, highlighting ResNet50 as the top-performing archi-tecture based on its robust accuracy and F1 score metrics. The experimental result depicts that transfer learning on pretrained ResNet50 architecture outperformed InceptionV3 and VGG19, achieving a test accuracy of 97% and F1 score of 98%.