Using DUCK-Net for polyp image segmentation.

Area of Science:

• Medical Imaging Analysis
• Computer Vision
• Artificial Intelligence

Background:

• Accurate polyp segmentation is crucial for colon cancer diagnosis.
• Existing methods often require large annotated datasets, limiting their application.

Purpose of the Study:

• To develop a novel supervised convolutional neural network (DUCK-Net) for accurate medical image segmentation.
• To evaluate DUCK-Net's performance on polyp segmentation tasks, especially with limited training data.

Main Methods:

• Designed an encoder-decoder architecture with residual downsampling and custom convolutional blocks.
• Employed data augmentation techniques to enhance the training dataset.
• Validated the model on benchmark polyp segmentation datasets: Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, and ETIS-LARIBPOLYPDB.

Main Results:

• DUCK-Net achieved state-of-the-art performance across multiple metrics (Dice, Jaccard, Precision, Recall, Accuracy).
• Demonstrated strong generalization capabilities, performing excellently even with limited training data.
• Showcased effectiveness in polyp segmentation within colonoscopy images.

Conclusions:

• DUCK-Net offers a versatile and effective solution for medical image segmentation, particularly for polyp detection.
• The model's ability to learn from small datasets addresses a significant challenge in medical AI.
• This work contributes to advancing automated polyp segmentation for improved colonoscopy analysis.