Fully Convolutional Architectures for Multiclass Segmentation in Chest Radiographs.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep neural networks (NNs) excel in image analysis, prompting applications in medical imaging.
• Automated segmentation of anatomical organs in chest radiographs (CXRs) presents unique challenges.

Purpose of the Study:

• To propose and investigate novel NN architectures for automated multiclass segmentation of lungs, clavicles, and heart in CXRs.
• To address challenges like overfitting, parameter reduction, and imbalanced data in CXR analysis.

Main Methods:

• Fusing recent convolutional network concepts adapted for CXR segmentation.
• Implementing an architecture with delayed subsampling, exponential linear units, and restrictive regularization.
• Utilizing a multiclass configuration trained on the Japanese Society of Radiological Technology database.

Main Results:

• The proposed architecture outperforms state-of-the-art methods for all organs.
• Achieved mean Jaccard overlap scores of 95% (lungs), 86.8% (clavicles), and 88.2% (heart).
• Outperformed human observers in segmenting lungs and heart.

Conclusions:

• The developed NN architecture offers high accuracy for multiclass organ segmentation in CXRs.
• This approach demonstrates potential to enhance diagnostic capabilities in medical imaging.
• The model's performance indicates a significant advancement over existing methods and human capabilities.