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

Models for colorectal polyp classification lose accuracy when applied to images from a different clinical center, and training on the low-frequency approximation subband of a discrete wavelet transform has been proposed as a remedy. We set out to test that hypothesis on Kvasir-SEG and PICCOLO, and report instead that three measurement practices common in this literature do not support the conclusions drawn from them. First, a single-level Haar LL subband is arithmetically identical to grayscale conversion followed by 2× decimation, verified to floating-point precision on both random and endoscopic data. At one decomposition level, wavelet decomposition and resolution reduction are not separable hypotheses, and the proposed mechanism is therefore not well-formed as posed. Second, image-level splitting—the standard remedy for patch-level leakage—is insufficient for video-derived datasets. PICCOLO's 3,433 frames derive from 70 physical lesions in 40 examinations; an image-level split leaves 58 of those lesions and 37 of those videos shared between training and test, and in-distribution AUC falls by 0.054 when measured against the partition the dataset authors supply. Third, the intersection-over-union metric used to assess attention localization inverts on high-coverage patches, where a polyp patch is almost entirely polyp: random attention scores 0.59 against 0.24 for trained heads. Measured without the area confound, localization is real but weak and confined to particular heads. We additionally quantify a reproducibility floor. In small models trained from scratch, runs receiving bit-identical inputs diverge by up to 0.177 AUC in the CNN/PICCOLO row through early-stopping checkpoint selection, and in one experimental arm this null control is itself statistically significant, so no claim from that arm is supportable at any p -value. A color-removal benefit observed in this regime (+0.067 AUC, 95% CI [+0.037, +0.097]) does not replicate on ImageNet-pretrained ResNet-18 and DeiT-Tiny backbones, where intervals exclude effects a fifth its size. We recommend reporting a null control alongside every comparison, and show that this identifies arms in which no claim can be supported.