Federated lung nodule segmentation using a hybrid transformer-U-Net architecture.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate lung cancer screening relies on early identification of malignant pulmonary nodules in CT scans.
• Segmentation is challenging due to nodule characteristics like small size, variable shape, low contrast, and class imbalance.

Purpose of the Study:

• To develop a privacy-preserving federated learning framework for accurate pulmonary nodule segmentation.
• To integrate a hybrid U-Net-Transformer architecture with a Dice-Focal loss function to overcome segmentation challenges.

Main Methods:

• A federated learning framework was designed using a hybrid U-Net-Transformer model.
• The model incorporates residual convolutional blocks and Transformer self-attention for feature extraction.
• A customized Dice-Focal loss function was employed to address class imbalance and improve boundary sensitivity.

Main Results:

• The federated framework achieved high segmentation fidelity for solid pulmonary nodules (15-25mm) on the LUNA16 dataset.
• Dice coefficients reached up to 0.93, demonstrating precise segmentation.
• The system showed robust precision-recall trade-offs across heterogeneous client datasets.

Conclusions:

• The proposed framework offers a scalable and privacy-preserving solution for lung cancer screening.
• The hybrid model and Dice-Focal loss effectively improve pulmonary nodule segmentation.
• Future work should focus on segmenting challenging nodules, such as part-solid or low-contrast types.