Pneumothorax Image Segmentation and Prediction with UNet++ and MSOF Strategy

Abstract

Deep learning is becoming more and more popular to solve image segmentation tasks in medical image processing community because of the incredible advantages in deep feature representation and nonlinear problem modeling. However, most existing deep learning methods based segmentation are implemented by combing deep, semantic, coarse-grained feature maps from the decoder sub network with shallow, low-level, fine-grained feature maps from the encoder sub-network, which are not up to the mustard of medical image segmentation. To solve the above-mentioned problem, an innovative end-to-end Pneumothorax Segmentation (PS) method based on UNet++ is proposed, where change maps could be learned from scratch using existing annotated datasets. And the fusion strategy of multiple side outputs is applied to combine change maps from different semantic levels. The high efficiency and availability of our proposed method are proved with SIIM-ACR Pneumothorax Segmentation dataset. Plenty of experimental results have shown that our proposed approach outperforms many cutting-edge methods.