MetaLung: Meticulous affine-transformation-based lung cancer augmentation method

Abstract

The limitation of medical image data in open source is a big challenge for medical image processing. Medical data is closed because of confidential and ethical issues, also manual labeling of medical data is an expensive process. We propose a new augmentation method named MetaLung (Meticulous affine-transformation-based lung cancer augmentation method) for lung CT image augmentation. The key feature of the proposed method is the ability to expand the training dataset while preserving clinical and instrumental features. MetaLung shows a stable increase in image segmentation quality for three CNN-based models with different computational complexity (U-Net, DeepLabV3, and MaskRCNN). Also, the method allows in reduce the number of False Positive predictions.