Study Finds Common Wavelet Preprocessing in Polyp Classification Is Simple Downsampling

A new evaluation of machine learning models designed for colorectal polyp classification indicates that widely used wavelet preprocessing does not function as previously believed and that standard evaluation practices fail to support claims of improved cross-dataset generalization. Researchers investigating cross-center performance evaluated models on the Kvasir-SEG and PICCOLO benchmarks to determine whether training on the low-frequency approximation subband of a discrete wavelet transform resolves clinical performance drops. Rather than validating the technique as a specialized feature extractor, the analysis demonstrated that several prevailing measurement methodologies produce misleading conclusions.

The study established that applying a single-level Haar low-low subband is mathematically equivalent to converting an image to grayscale and decimating it by a factor of two. The investigators verified this arithmetic identity to floating-point precision across both synthetic random data and clinical endoscopic imagery. Because single-level wavelet decomposition and spatial resolution reduction represent identical operations in this setting, the authors concluded that wavelet filtering and downsampling are not separable hypotheses, rendering the proposed theoretical mechanism ill-formed as originally posed.

The investigation also revealed significant data leakage issues arising from standard data-splitting methodologies applied to endoscopic video datasets. While image-level splitting is commonly treated as an adequate defense against patch-level data leakage, the authors demonstrated that it remains insufficient for video-derived benchmarks. In the PICCOLO dataset, which contains 3,433 frames captured from 70 physical lesions across 40 clinical examinations, a conventional image-level partition caused 58 physical lesions and 37 videos to be simultaneously shared between the training and test sets.

When model evaluation accounted for this video-level overlap by comparing results against the patient-level partition supplied by the dataset creators, the in-distribution area under the receiver operating characteristic curve decreased by 0.054.

The authors additionally examined attention interpretability, discovering that the intersection-over-union metric frequently used to quantify attention localization fails on high-coverage image patches. When a patch consists almost entirely of polyp tissue, the metric inverted, yielding an intersection-over-union score of 0.59 for completely random attention maps compared to only 0.24 for trained attention heads. Once this patch area confound was removed from the evaluation, true model localization was observed to be real but weak, and strictly confined to specific individual attention heads.

Beyond metric failures, the authors quantified a substantial reproducibility floor among small neural networks trained entirely from scratch. When identical convolutional neural networks received bit-identical inputs on the PICCOLO dataset, independent training runs diverged by as much as 0.177 in area under the receiver operating characteristic curve solely due to early-stopping checkpoint selection. In one experimental arm, the variation of this null control reached statistical significance, demonstrating that no substantive claim could be supported at any p-value in that setting.

This instability directly affected apparent architectural improvements, such as an observed benefit from color removal. While color removal appeared to yield an advantage of 0.067 area under the curve with a 95 percent confidence interval between 0.037 and 0.097 in small scratch-trained models, this effect failed to replicate on ImageNet-pretrained ResNet-18 and DeiT-Tiny backbones, where confidence intervals excluded effects even one-fifth of that magnitude. Based on these findings, the authors recommended that machine learning studies consistently report null controls alongside experimental comparisons to identify evaluation arms where variations exceed the limits of statistical support.

Similarly, the team demonstrated that patient-level data partitioning is critical when evaluating models derived from endoscopic video footage.

Subject of Research: Evaluation of wavelet preprocessing and validation methodology in machine learning for colorectal polyp classification

Article Title: Single-level wavelet preprocessing is downsampling: measurement failures in cross-dataset polyp classification

Article References: R., J., Tewari, D., Veerapu, G., J. S., N., Palanisamy, G., & Kohli, S. (2026). Single-level wavelet preprocessing is downsampling: measurement failures in cross-dataset polyp classification. Frontiers in Artificial Intelligence, 9, Article 1929384. https://doi.org/10.3389/frai.2026.1929384

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1929384

Keywords: colorectal polyp classification, wavelet preprocessing, discrete wavelet transform, Haar wavelet, data leakage, patient-level splitting, cross-dataset generalization, PICCOLO dataset, Kvasir-SEG, attention interpretability, intersection-over-union, reproducibility, early stopping, convolutional neural networks, ResNet-18, DeiT-Tiny