A Bi-FPN-Based Encoder-Decoder Model for Lung Nodule Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Accurate lung nodule segmentation in CT images is crucial for early lung cancer detection.
• Challenges include nodule variability in shape, appearance, and surrounding tissues.
• Existing methods often lack robustness and efficiency.

Purpose of the Study:

• To propose a resource-efficient, end-to-end deep learning model for robust lung nodule segmentation.
• To enhance segmentation accuracy using a Bi-FPN architecture and Mish activation.
• To improve model performance through weighted binary cross-entropy loss.

Main Methods:

• Developed an end-to-end deep learning model integrating a Bi-FPN between encoder and decoder.
• Utilized the Mish activation function and class weights for enhanced segmentation efficiency.
• Trained and evaluated the model on the LUNA-16 and QIN Lung CT datasets.
• Employed weighted binary cross-entropy loss for improved voxel classification.

Main Results:

• The proposed model achieved a Dice Similarity Coefficient of 82.82% on LUNA-16 and 81.66% on the QIN dataset.
• Demonstrated superior performance compared to existing deep learning models like U-Net.
• Indicated robustness and efficiency in lung nodule segmentation.

Conclusions:

• The developed deep learning architecture offers an efficient and accurate solution for lung nodule segmentation.
• This approach holds significant potential for improving early lung cancer diagnosis.
• The model's performance highlights the effectiveness of Bi-FPN and weighted loss functions.