Extracting Lungs from CT Images via Deep Convolutional Neural Network Based Segmentation and Two-Pass Contour...

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Artificial Intelligence in Medicine

Background:

• Lung segmentation is crucial for thoracic CT image analysis and computer-aided pulmonary disease diagnostics.
• Challenges include image noise, pathologies, vessels, and anatomical variations, complicating accurate segmentation.
• Existing methods often struggle with these complexities, necessitating improved automated approaches.

Purpose of the Study:

• To develop a fully automatic and accurate algorithm for lung segmentation in thoracic CT images.
• To address the limitations of current lung segmentation techniques, particularly in the presence of complex thoracic structures and pathologies.
• To enhance the reliability of lung segmentation for downstream computer-aided diagnostic tasks.

Main Methods:

• An input image is divided into fixed-sized patches for initial lung region extraction using a deep convolutional neural network.
• Superpixel segmentation is applied, followed by local contour refinement using an adjacent point statistics method.
• Global contour refinement is achieved through edge direction tracing to include juxta-pleural lesions.

Main Results:

• The algorithm achieved an average Dice similarity coefficient of 97.95% and Jaccard's similarity index of 94.48%.
• Low average segmentation errors were observed: 2.8% over-segmentation and 3.3% under-segmentation compared to manual segmentation.
• The method demonstrated superior performance compared to several feature-based machine learning and current lung segmentation techniques.

Conclusions:

• The proposed fully automatic algorithm provides accurate lung segmentation in thoracic CT images.
• The method effectively handles complexities like noise and pathologies, improving upon existing techniques.
• This advancement holds significant potential for enhancing computer-aided pulmonary disease diagnostics.