FFTMed: leveraging fast-fourier transform for a lightweight and adversarial-resilient medical image segmentation...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate medical image segmentation is crucial for computer-aided diagnosis and treatment planning.
• Image noise significantly degrades diagnostic accuracy and complicates treatment.
• Existing methods partially integrate frequency-domain processing, limiting their effectiveness.

Purpose of the Study:

• To introduce FFTMed, a novel neural architecture for medical image segmentation operating directly in the frequency domain.
• To enhance noise resilience and computational efficiency in medical image segmentation.
• To propose a new benchmark for evaluating noise susceptibility in segmentation models.

Main Methods:

• Developed FFTMed, a neural network architecture that processes images in the frequency domain.
• Utilized the inherent noise resilience of frequency-domain processing.
• Introduced a new benchmark dataset with varying noise levels for rigorous evaluation.

Main Results:

• FFTMed effectively eliminates noise and achieves accurate medical image segmentation.
• The model demonstrates resilience to noisy test images without additional noise augmentation during training.
• FFTMed exhibits robust resistance to adversarial attacks compared to baseline models.

Conclusions:

• FFTMed offers a powerful solution for noise reduction and accurate segmentation in medical imaging.
• The proposed architecture improves diagnostic accuracy and treatment planning efficacy.
• FFTMed provides a robust and computationally efficient approach for real-world medical image analysis.