Automatic ۳D Analysis of Bronchial Tree Dilatation Using Deep Learning Algorithms in Chest CT images

Background and Aims: Precise segmentation of the bronchial airway tree in chest CT images is vital for early diagnosis, treatment planning, and image-guided procedures in pulmonary diseases such as COPD and lung cancer. However, due to the low contrast, thin walls, and complex branching patterns of peripheral airways, existing techniques often struggle with accuracy. This study proposes a novel deep learning framework designed to enhance segmentation performance, particularly in detecting fine airway structures. Methods: We introduce ۳D AIR-Net (Airway-Informed Refinement Network), a novel ۳D CNN architecture integrating multi-scale feature extraction and a Boundary Enhancement Module (BEM). This design allows for the preservation of thin airway walls and accurate delineation of complex branches. The model was trained and validated on two publicly available datasets—ATM’۲۲ and EXACT’۰۹—after preprocessing steps such as HU normalization, volume resampling, and patch extraction with ۵۰% overlap. A combination of Dice and Tversky loss was used to address class imbalance and improve the detection of peripheral airways. Results: ۳D AIR-Net achieved a Dice Similarity Coefficient of ۰.۸۸۲, a precision of ۰.۸۶۱, and a Tree Length Detected Rate (TD) of ۰.۸۴۵, outperforming various state-of-the-art models. Visual assessments confirmed the model’s effectiveness in preserving airway continuity, enhancing boundary sharpness, and minimizing false positives, especially in peripheral regions. Conclusion: The ۳D AIR-Net framework demonstrates high accuracy in ۳D airway segmentation and holds strong potential for clinical applications such as bronchoscopy navigation and pre-operative planning. Its capacity to detect fine, complex airway structures makes it a valuable tool for improving diagnostic and interventional outcomes. Further studies will focus on its performance in diverse imaging conditions and pathological scenarios.

کلیدواژه ها:

Airway Segmentation ، Deep Learning ، Chest CT ، ۳D CNN ، Pulmonary Imaging ، Artificial Intelligence