Inter-Class Analysis of Frequency-Band Similarity in Gastrointestinal Endoscopic Image Datasets

Abstract

Gastrointestinal (GI) endoscopic examinations can detect various GI issues early. Challenges like high intra-class variability, moderate differences between classes, and biased data complicate automated classification. The study analyzes frequency-dependent image features in classification results, focusing on similarities within and between classes to better understand the dataset and identify the most difficult classes to classify. It uses a similar approach for similarity analysis with Discrete Wavelet Transform (DWT), breaking images into low-frequency (LL) and high-frequency (HH) sub-bands based on frequency ranges. Structural Similarity Index Measure (SSIM) and Mean Squared Error (MSE) inter- and intra- class similarities. Functional validation involved a classification test using the Random Forest (RF) model. Experiments on multiple GI endoscopic datasets illustrate that LL sub-bands, capturing coarse structural features, provide higher discriminative power and improve classification accuracy, while HH sub-bands, preserving fine textures, are less effective due to higher inter-class similarity. Analysis of similarity measures highlights classes with high intra-class variability, particularly minority classes, as the most challenging for classification. The frequency-aware similarity approach enhances interpretability, reveals dataset-specific issues, and automates the evaluation of gastrointestinal images.