# Lung Structures Enhancement in Chest Radiographs via CT based FCNN Training

**Authors:** Ophir Gozes, Hayit Greenspan  
**arXiv:** [1810.05989](https://arxiv.org/abs/1810.05989)  
**Date:** 2018-10-14  
**Summary page:** https://www.summarizepaper.com/en/paper/1810.05989v1/

## Key points

- Deep learning-based approach for enhancing soft lung structures in chest radiographs
- Technique using Fully Convolutional Neural Networks (FCNN) to improve contrast of soft lung structures
- Training two 2D FCNN architectures: one for lung segmentation and another for extracting lung structures
- Training images created using Simulated X-Ray or Digitally Reconstructed Radiographs (DRR)
- Dataset of 2D lung masks created for training the segmentation FCNN
- DRR images used as target images for training the extraction FCNN
- Lung structures can be enhanced by fusing them with original input x-ray
- Applicability of enhancement technique to real x-ray data demonstrated on NIH Chest X-Ray-14 dataset
- Promising results when training a DenseNet-121 based architecture directly on the lung-enhanced X-ray images
- Potential for improving automated algorithms' performance in detecting lung pathology

## Summary

This paper proposes a deep learning-based approach for enhancing soft lung structures in chest radiographs by addressing overlapping anatomical structures. It introduces a technique using Fully Convolutional Neural Networks (FCNN) to improve the contrast of soft lung structures. The technique involves training two 2D FCNN architectures: one for 2D lung segmentation to normalize the lung area and another for extracting lung structures. Training images are created using Simulated X-Ray or Digitally Reconstructed Radiographs (DRR) derived from scans belonging to the LIDC-IDRI dataset. A dataset of 2D lung masks is created for training the segmentation FCNN and DRR images containing voxels belonging to the 3D lung segmentation are used as target images for training the extraction FCNN. Once the lung structures are extracted, they can be enhanced by fusing them with the original input x-ray. The authors demonstrate that their enhancement technique is applicable to real x-ray data and present results on the NIH Chest X-Ray-14 dataset. They also highlight promising results when training a DenseNet-121 based architecture directly on the lung-enhanced X-ray images. In conclusion, this technique shows potential for improving automated algorithms' performance in detecting lung pathology and can be applied to real x-ray data with promising results.