Lung segmentation on standard and mobile chest radiographs using oriented Gaussian derivatives filter

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Unsupervised lung segmentation is essential for developing Content-Based Medical Image Retrieval Systems (CBMIRS) for chest radiographs (CXRs).
• Existing methods may lack robustness or require manual intervention, necessitating automated solutions.

Purpose of the Study:

• To present a robust, fully automated, unsupervised method for lung segmentation of standard and mobile chest radiographs.
• To enhance the development of CBMIRS for CXR analysis.

Main Methods:

• Utilized oriented Gaussian derivatives filters (seven orientations) combined with Fuzzy C-Means (FCM) clustering and thresholding.
• Developed a novel algorithm for automatic threshold value generation for Gaussian responses.
• Applied pre-processing blocks to standardize images from various machines (PA and AP views) and datasets (JSRT, private).

Main Results:

• Achieved high performance measures (accuracy, F-score, precision, sensitivity, specificity > 0.90) on the JSRT dataset, with overlap measure at 0.87.
• Demonstrated robust performance on private datasets: overlap measure of 0.81 (standard) and 0.69 (mobile).
• The algorithm is fully automated and fast, averaging 12.5 seconds for 512x512 images.

Conclusions:

• The proposed method is fully automated and unsupervised, requiring no training or learning stages for lung segmentation.
• Pre-processing blocks effectively standardize radiographs from mobile machines.
• The algorithm offers good performance, robustness, and speed, suitable for CBMIRS applications.