Efficient colorectal polyp segmentation using wavelet transformation and AdaptUNet: A hybrid U-Net.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Colorectal cancer often develops from polyps, necessitating early detection via colonoscopy.
• Variability in polyp appearance complicates accurate identification in medical images.
• Existing methods face challenges in precise polyp segmentation.

Purpose of the Study:

• To develop an advanced deep learning model for accurate polyp segmentation in colonoscopy images.
• To enhance the early detection of colorectal polyps, aiding in cancer prevention.
• To improve the efficiency and precision of automated polyp detection systems.

Main Methods:

• A customized U-Net deep learning architecture, named AdaptUNet, was developed.
• Attention mechanisms and skip connections were integrated for detailed feature analysis.
• Wavelet transformations were employed to extract subtle, often overlooked image features.

Main Results:

• AdaptUNet achieved a Dice coefficient of 0.9104 and IoU of 0.8368 on the CVC-300 dataset.
• The model demonstrated high Balanced Accuracy (0.9880) and strong performance across multiple datasets (Kvasir-SEG, Etis-LaribDB).
• Training on the Hyper Kvasir dataset confirmed the model's robustness with diverse data.

Conclusions:

• The proposed AdaptUNet model offers an efficient and high-performance solution for colorectal polyp detection.
• This deep learning approach promises improved accuracy and reduced manual effort in colonoscopy analysis.
• The findings suggest a significant advancement in automated diagnostic tools for colorectal cancer screening.