AI models for colorectal-polyp images can appear stronger when training and testing share information from the same lesions.
New experiments exposed this problem alongside weaknesses in common image-processing comparisons.
What researchers studied
Researchers based in India examined classification experiments using the Kvasir-SEG and PICCOLO endoscopy datasets.
They investigated preprocessing, data splitting, attention-map evaluation and reproducibility, rather than conducting a clinical screening trial.
PICCOLO contains 3,433 frames from 70 physical lesions in 40 examinations.
Frames from the same lesion can be closely related, so splitting individual images does not necessarily create independent training and test sets.
The team compared evaluation approaches and repeated computational experiments.
What they found
An image-level split placed 58 lesions and 37 videos on both sides of the training-test boundary.
Compared with the dataset authors’ supplied partition, measured in-distribution area under the curve fell by 0.054.
The researchers also demonstrated that single-level Haar wavelet preprocessing is mathematically equivalent to grayscale conversion followed by halving resolution.
Some attention-localisation measures were misleading, and identical-input training runs diverged by up to 0.177 AUC in one small-model condition.
What it means
A reported benefit from removing colour did not replicate on the tested pretrained backbones.
Together, these findings challenge claims that a more elaborate image-processing method necessarily improves generalisation.
Developers should split datasets at the level of independent lesions or examinations when appropriate, test against simple preprocessing baselines and report reproducibility controls.
These precautions matter beyond colonoscopy wherever medical datasets contain multiple related observations.
Important limitation
Apparent model accuracy is only as credible as the evaluation design. A high score cannot establish real-world clinical performance when test observations are not sufficiently independent.
The study involved specific public datasets and selected architectures, not prospective deployment in hospitals.
It does not establish patient outcomes or invalidate every polyp classifier.
Its central contribution is identifying evaluation practices that can mislead.
Source Information
Study: Single-level wavelet preprocessing is downsampling: measurement failures in cross-dataset polyp classification
Authors: Jenifer R.; Dhimant Tewari; Goutham Veerapu; Nisha J. S.; Gopinath Palanisamy; Saanji Kohli
Journal / Source: Frontiers in Artificial Intelligence
Published: 9 November 2026
