Table 1.
DL based methods for segmentation of CXR images.
| Approaches | Strategy | Dataset & Results | Pros/Cons |
|---|---|---|---|
| Novikov et al. [5] | Modified U- Net | JSRT CXR dataset | Improvement in experimental results can be achieved using more dataset |
| Overlap: lungs 95%, heart 88.2% and clavicles 86.8% | |||
| Wang [6] | Multi-task Fully Convolutional Networks | JSRT CXR dataset | More dataset can be used to improve the performance of the model |
| Overlap: Lungs 96% heart 94.3% and clavicles 90.6% | |||
| Arbabshirani et al. [7] | Multi-scale CNN | PACS and JSRT dataset | The patch-based approach applied but performance is low in comparison to other methods |
| Overlap: 91% | |||
| Jiang et al. [8] | CNN model with Recurrent layers | JSRT Dataset | Only quantitative results are given |
| Quantitative results not given | |||
| Xie et al. [9] | Multi-view knowledge-based collaborative (MV-KBC) model using ResNet-50 and U-Net model | LIDC-IDRI dataset | Low-performance results and Patch-based approach applied |
| DSC: 80.23% | |||
| Eslami et al. [10] | Mukti-task organ segmentation using conditional Generative Adversarial Network (GAN) pix2pix network | JSRT dataset | Outperforms other multi-class segmentation methods |
| Overlap:98.4%, | |||
| DSC: 99.2% | |||
| Park et al. [11] | Deep CNN model for segmentation of Diffuse Interstitial Lung Diseases (DILD) | High-Resolution Computed Tomography (HRCT) dataset | The model is applied to different diseases. |
| Overlap: 96.76%, | |||
| DSC: 98.84% | |||
| Gaál et al. [12] | Adversarial Attention U-Net model for lung segmentation | JSRT and MC dataset | Good performance for lung segmentation |
| DSC: 97.5% | |||
| Wang et al. [13] | Multitask Dense U-Net (MDU-Net) model for rib and clavicle segmentation | CXR dataset | A small dataset of only 88 CXR images used, no data extension method is used |
| DSC: 88.38% |