Automatic lung parenchyma segmentation using a deep convolutional neural network from chest X-rays. (March 2022)

• Record Type: Journal Article
• Authors: Maity, Arunit; Nair, Tusshaar R.; Mehta, Shaanvi; Prakasam, P.
• Is Part Of: Biomedical signal processing and control. Volume 73(2022)
• Publication Date: 2022-03
• Subjects: Deep convolutional neural network -- Chest X-ray -- Lung parenchyma segmentation -- Jaccard similarity coefficient -- Dice similarity coefficient -- Ablation study; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.103398
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Proposed the DCNN model for an automatic lung segmentation from CXR images. Incorporated the residual connections in the proposed DCNN model to increase the efficiency. Proposed the Penalty Combo Loss (PCL) function which helps us to identify the region as lung even when there's no lung present and vice versa. The proposed deep convolutional neural network is trained and tested with JSRT NLM MC and Shenzhen dataset. This method achieved a Dice Similarity Coefficient of 0.982 ± 0.018 and a Jaccard Similarity Coefficient of 0.967 ± 0.015 on lung segmentation in chest X-rays. Abstract: To detect and diagnosis the lungs related diseases, a Chest X-Ray (CXR) is the major tool used by the physician. Automated organ segmentation contributes to a crucial part of computer-aided detection (CAD) and diagnosis of diseases from CXRs as they enhance the accuracy in detecting, and also help in reducing the burden raised due to manual diagnosis from radiologists and medical practitioners. In this paper, an efficient automatic CAD system is proposed to detect the boundaries using a deep convolutional neural network (DCNN) model. The DCNN is trained in an end-to-end setting to facilitate fully automatic lung segmentation from anteroposterior or posteroanterior view CXRs. It learns to predict binary masks for a given CXR, by learning to discriminate regions of organ parenchyma from regions of no organ parenchyma. The proposed model's architecture makes use of residual connections in all the concurrent up-sampling paths from each encoder block at every level, thus facilitating collective learning within blocks through inter-sharing of all high-dimensional feature maps. To generalize the proposed model to CXRs from all data distributions, image preprocessing techniques such as Top-Hat Bottom-Hat Transform and Contrast Limited Adaptive Histogram Equalization are employed. The proposed model is trained and tested using the JSRT, NLM-MC and Shenzhen Hospital datasets. The proposed method achieved a Dice Similarity Coefficient of 0.982 ± 0.018 and a Jaccard Similarity Coefficient of 0.967 ± 0.015. The implementation results demonstrated that the proposed method has surpassed the existing methods and our model is relatively lightweight and can be easily implemented on standard GPUs.