Nested Attention Network for Robust Medical Image Segmentation Under Digital Watermarking

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Digital Security

Background:

• Digital watermarking secures medical images but may affect AI analysis.
• Watermarking's impact on image classification is known, but its effect on segmentation is less understood.

Purpose of the Study:

• To investigate the effects of digital watermarking on medical image segmentation performance.
• To evaluate how watermarking influences deep learning-based segmentation models.
• To assess a novel deep learning model designed for enhanced watermarked image analysis.

Main Methods:

• Evaluated three watermarking techniques on U-Net, ResUNet++, SegNet, FCDenseNet, and TernausNet models.
• Tested models on LIDC-IDRI and BRISC medical image datasets.
• Developed and assessed a novel deep learning model with nested attention mechanisms.

Main Results:

• Digital watermarking caused minor performance degradation in segmentation models across datasets.
• Mean Intersection over Union (mIoU) reduction was between 0.15%-0.44% (BRISC) and 0.19%-0.29% (LIDC-IDRI).
• The novel attention-based model showed improved sensitivity to subtle variations in watermarked images.

Conclusions:

• Digital watermarking is compatible with AI-based medical image segmentation.
• Watermarking techniques offer a viable solution for protecting medical images without significantly compromising segmentation accuracy.
• Findings support the broader clinical application of watermarked medical images in AI systems.