Structure Correction for Robust Volume Segmentation in Presence of Tumors

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Artificial Intelligence

Background:

• Convolutional Neural Network (CNN) based lung segmentation models struggle with severe pathologies like tumors due to limited diverse training data.
• Accurate lung segmentation is crucial for diagnosing and monitoring various lung conditions.

Purpose of the Study:

• To develop a robust multi-stage algorithm for precise lung volume segmentation from CT scans, particularly in the presence of significant pathologies.
• To improve the segmentation accuracy and recall of pathological lung features, such as tumors, without compromising performance on normal lung tissue.

Main Methods:

• A multi-stage approach employing a 3D CNN for initial coarse segmentation, followed by a 3D structure correction CNN for mask refinement.
• Incorporation of a novel data augmentation strategy to train the 3D CNN, integrating global shape priors.
• Final segmentation refinement using a parallel flood-fill operation.

Main Results:

• The proposed algorithm demonstrates robustness in segmenting lungs with large nodules/tumors, improving recall of juxtapleural tumor voxels by at least 15% compared to state-of-the-art methods.
• Segmentation accuracy for normal lungs is maintained.
• The method achieves segmentation within 5 seconds, meeting Computer-Aided Diagnosis (CAD) software requirements.

Conclusions:

• The developed multi-stage algorithm offers a significant advancement in lung segmentation for CT scans, effectively handling severe pathologies.
• The approach provides a faster and more accurate solution for lung segmentation, beneficial for clinical applications and CAD systems.
• No labeled segmentation masks for entire pathological lung volumes are required for training, simplifying the process.