Comparison of Different Convolutional Neural Networks Utilizing Transfer Learning for Pneumothorax Segmentation from Whole Chest X-Ray Images and Extracted Patches

Pneumothorax is a lung condition characterized by the presence of air between chest wall and lungs. In order to diagnose location and size of pneumothorax, chest X-ray is a commonly used imaging technique. U-Net convolutional neural network models with different backbones are compared in order to assess their capability to automatically and correctly segment signs of pneumothorax from chest X-rays. Five different pretrained backbones have been chosen: VGG19, ResNet34, ResNet50, DenseNet121 and Inceptionv3. Two different approaches for pneumothorax segmentation have also been tested: one methodology used X-ray images of the whole chest area for training, while the second one split the original images into patches and used them for the training process. Both methodologies performed at a similar level, with the best results achieved by U-Net model with DenseNet121 backbone for segmentation of X-ray of the whole chest. This model achieved a Jaccard index and Dice score of 76.92% and 78.81%, respectively. These results indicate that the tested models are capable of extracting fine-grained features from X-ray images of whole chest and that patch-based segmentation does not provide additional benefits.