Introducing a secondary segmentation to construct a radiomics model for pulmonary tuberculosis cavities

Area of Science:

• Radiomics
• Medical Imaging
• Pulmonary Medicine

Background:

• Accurate lung segmentation is crucial for radiomic studies in non-neoplastic lung diseases like pulmonary tuberculosis (PTB).
• Traditional methods require precise delineation of diseased areas, posing a challenge for automated analysis.

Purpose of the Study:

• To develop and validate a novel segmentation method for chest X-rays (CXR) to facilitate radiomic analysis of PTB.
• To construct radiomic models for automatic PTB cavity classification without manual disease delineation.

Main Methods:

• A retrospective study utilized 266 PTB patient CXRs.
• A U-net-based model performed primary lung segmentation, followed by a secondary sliding window segmentation.
• Pyradiomics extracted features, which were consolidated using standard deviation and variance, followed by dimensionality reduction and Random Forest model construction.

Main Results:

• Two radiomic signatures with 10 texture features each were identified.
• Random Forest models demonstrated high performance, with the standard-deviation model achieving an AUC of 0.9444 and the variance model an AUC of 0.9288.
• The models accurately classified PTB cavities from normal CXRs.

Conclusions:

• The developed secondary sliding window segmentation method eliminates the need for precise disease delineation in pulmonary radiomic studies.
• This approach enhances the accuracy of CXR reporting, supporting its role in high-volume screening for PTB.
• The radiomic models effectively classify PTB cavities, improving diagnostic capabilities.