Multiclass Convolution Neural Network for Classification of COVID-19 CT Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Diseases

Background:

• The novel coronavirus (COVID-19) pandemic presents diagnostic challenges due to symptom overlap with other respiratory illnesses.
• Accurate and rapid classification of COVID-19 is crucial for effective patient management and public health.
• Computed Tomography (CT) imaging shows promise for COVID-19 diagnosis, but automated analysis requires robust models.

Purpose of the Study:

• To evaluate the effectiveness of transfer learning with deterministic algorithms for classifying COVID-19 using CT images.
• To compare the performance of various Convolutional Neural Network (CNN) architectures in identifying COVID-19 positive and negative cases.
• To identify the optimal CNN models for accurate COVID-19 detection from CT scans.

Main Methods:

• A dataset of 746 CT images (COVID-19 and non-COVID-19) was utilized, split into training, validation, and testing sets.
• Data augmentation techniques were applied to enhance the training dataset size.
• Transfer learning with deterministic algorithms was implemented, and CNN models (ResNeXt101, ResNet152, GoogleNet, DenseNet201, ResNet101) were pretrained for binary classification.

Main Results:

• ResNeXt101 achieved the highest F1 score (0.978) and accuracy (97.81%), along with 95.71% sensitivity and 100% specificity.
• ResNet152 demonstrated strong performance with an F1 score of 0.938, accuracy of 93.80%, 90% sensitivity, and 98.33% specificity.
• GoogleNet yielded a lower F1 score of 0.762, indicating varying performance across different CNN architectures.

Conclusions:

• Transfer learning combined with deterministic algorithms and CNNs offers a highly effective approach for COVID-19 classification from CT images.
• Models like ResNeXt101 and ResNet152 show significant potential for accurate and reliable COVID-19 detection in clinical settings.
• The study highlights the importance of selecting appropriate CNN architectures for optimal diagnostic performance in medical image analysis.