A Transfer Learning Method for Covid-19 and Pneumonia Diagnosis Based on Chest Radiograph Classification

Pneumonia has been a tough and dangerous human illness for a history-long time, notably since the COVID-19 pandemic outbreak. Many pathogens, including bacteria or viruses like COVID-19, can cause pneumonia, leading to inflammation in patients' alveoli. A corresponding symptom is the appearance of lung opacities, which are vague white clouds in the lungs' darkness in chest radiographs. Modern medicine has indicated that pneumonia-associated opacities are distinguishable and can be seen as fine-grained labels, which make it possible to use deep learning to classify chest radiographs as a supplementary aid for disease diagnosis and performing pre-screening. However, deep learning-based medical imaging solutions, including convolutional neural networks, often encounter a performance bottleneck when encountering a new disease due to the dataset's limited size or class imbalance. This study proposes a deep learning-based approach using transfer learning and weighted loss to overcome this problem. The contributions of it are three-fold. First, we propose an image classification model based on pre-trained Densely Connected Convolutional Networks using Weighted Cross Entropy. Second, we test the effect of masking non-lung regions on the classification performance of chest radiographs. Finally, we summarize a generic practical paradigm for medical image classification based on transfer learning. Using our method, we demonstrate that pre-training on the COVID-19 dataset effectively improves the model's performance on the non-COVID Pneumonia dataset. Overall, the proposed model achieves excellent performance with 95.75% testing accuracy on a multiclass classification for the COVID-19 dataset and 98.29% on a binary classification for the Pneumonia dataset.