Fully automatic deep learning-based lung parenchyma segmentation and boundary correction in thoracic CT scans.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Deep Learning

Background:

• Accurate lung parenchyma segmentation is crucial for diagnosing respiratory diseases.
• Existing methods may struggle with juxta-pleural nodules and segmentation inaccuracies.

Purpose of the Study:

• To develop a precise lung parenchyma segmentation algorithm for thoracic CT scans.
• To combine deep learning with traditional image processing for improved segmentation.
• To correct segmentation errors, including false positives and negatives.

Main Methods:

• An improved 2D U-Net convolutional neural network (CNN) with Inception-ResNet-v2 backbone was trained.
• The model was trained on 32 CT scans and evaluated on 16 scans with juxta-pleural nodules.
• A post-processing algorithm was implemented to refine lung masks and include juxta-pleural nodules.

Main Results:

• The CNN model achieved high performance metrics (DSCavg: 0.9791, IoUavg: 0.9624, F1avg: 0.9792).
• Post-processing further refined segmentation (DSCavg: 0.9713, IoUavg: 0.9486, F1avg: 0.9701).
• The post-processing successfully incorporated juxta-pleural nodules into the final lung masks.

Conclusions:

• The proposed CNN-based method achieves precise lung parenchyma segmentation.
• The post-processing algorithm effectively addresses segmentation errors.
• This approach holds potential for advancing automated nodule detection in CAD systems.