Adaptive Frequency Learning Network With Anti-Aliasing Complex Convolutions for Colon Diseases Subtypes.

Area of Science:

• Medical Imaging
• Computational Pathology
• Gastroenterology

Background:

• Accurate colon disease subtyping via colonoscopy is vital for effective staging and treatment.
• Inter-class confusion and brightness variations hinder current subtyping methods.
• Fourier-based image spectra offer potential due to frequency features and brightness insensitivity.

Purpose of the Study:

• To develop a robust framework for colon disease subtyping using frequency domain learning.
• To address challenges of inter-class confusion and brightness imbalance in endoscopic images.
• To leverage Fourier spectra for improved accuracy in colon disease classification.

Main Methods:

• A novel frequency domain learning framework incorporating four core modules: position consistency, high-frequency self-supervised learning, complex number arithmetic, and feature anti-aliasing.
• Position consistency module preserves spectral information and enhances training stability.
• High-frequency autoencoder guides selective learning of relevant frequency features.
• Complex number arithmetic model enables direct spectral training without phase information loss.
• Feature anti-aliasing module prevents spectral aliasing during down-sampling.

Main Results:

• The proposed framework achieved state-of-the-art performance on a five-class dataset of 4591 colorectal endoscopic images.
• An accuracy rate of 89.82% was obtained, demonstrating the method's effectiveness.
• The integrated modules successfully addressed spectral data challenges like brightness imbalance and inter-class confusion.

Conclusions:

• The developed frequency domain learning framework significantly improves the accuracy of colon disease subtyping.
• This approach offers a promising solution for reliable automated analysis of colorectal endoscopic images.
• The method facilitates more in-depth disease staging and personalized treatment planning.