More frames or better labels?: training data and computer-aided detection of sessile serrated lesions

In this issue of Clinical Endoscopy, Han et al.1 tackle a deceptively simple question: With an algorithm held constant, can better training data improve the computer-aided detection (CADe) of sessile serrated lesions (SSLs)? This question was clinically addressed. A prior meta-analysis of 28 randomized trials found that CADe raised the adenoma detection rate (ADR) by 20% and more than halved the adenoma miss rate, but did not alter the SSL detection rate (risk ratio, 1.10; 95% confidence interval [CI], 0.93–1.30).2 This omission matters, because the proximal serrated polyp detection rate is inversely associated with interval post-colonoscopy colorectal cancer, independent of the ADR.3 Improving SSL detection is therefore not simply another technical benchmark; it tests whether artificial intelligence (AI) can help where human visual recognition is most vulnerable.

The investigators trained two models with identical Visual Geometry Group 16 architecture, hyperparameters, and an inference pipeline, varying only the input data. Model A used approximately 54,000 frames from public datasets labeled by visual consensus, while model B used approximately 120,000 frames (more than twice as many) from 431 full-length colonoscopy videos recorded by two endoscopists with sustained ADRs above 35%, every lesion label matched one-to-one with a histopathology report. Volume and label provenance, therefore, moved together, a confound I wish to revisit. In 31 validation videos from a third, independent endoscopist, model B achieved an event-based sensitivity of 99.7% versus 39.9%, detected every adenoma and SSL event, halved the median detection latency, and held the lesions across more frames. The precision decreased from 96.0% to 79.0%, while false-positive events rose from 0.16 to 2.55 per video, although most alerts were brief.

This study’s most valuable contribution is conceptual rather than numerical. The training set was part of the intervention. A model does not learn “SSL” as a biological entity; it learns examples and labels and exclusions its developers supply. For serrated lesions, this is not a simple matter of degree. Separating an SSL from a hyperplastic polyp is a histological judgment,4 meaning a corpus labeled by visual consensus does not merely carry noise5—it may never encode the category at all, which remains a ceiling no architecture can lift. A full-procedure video also exposes the model to folds, bubbles, debris, instruments, and unremarkable mucosa, which are important limitations that curated still image collections largely omit. Meanwhile, most public colonoscopy databases under-report lesions, patient characteristics, histology, and the annotation process itself.6

In this context, quieter observations deserve further attention. Although hyperplastic polyps were more common at the lesion level, neoplastic lesions dominated at the frame level as endoscopists generally wash, angulate, and inspect for longer when they suspect neoplasia. Careful practice thus enriches the corpus with no deliberate curation effort; routine high-quality colonoscopy behaves as an implicit annotation process and an argument for capturing videos from high detectors rather than simply more of it.

However, the study’s principal limitation deserves to be stated without hedging: the comparison is insufficient to isolate data quality as a causal factor. Model B was trained on more than twice as many frames as model A, which means that an unknown share of the performance gap is attributable to data volume alone, rather than curation, as dataset size and class distribution both independently affect performance in endoscopic deep learning.7 The two corpora also differed in terms of the source, lesion distribution, acquisition environment, and examiner expertise. Fixing the architecture controls one variable, but cannot separate the others, and the contrast between 99.7% and 39.9% should not be read as a measure of the contribution of pathology confirmation alone. The next step should involve prespecified ablation: size-matched subsets, learning curves, and pathology-confirmed versus visually assigned labels on the same videos. Model A also performed well below the standard of contemporary commercial CADe systems and represents a floor, rather than a state-of-the-art comparator.

Validation requires a similar calibration. Thirty-one videos, 14 unique SSLs, one center, and one endoscopist yield a compelling signal, rather than a stable estimate, and the 40 SSL events clustered within those 14 lesions; therefore, further cluster-aware analysis will be necessary. Event-based sensitivity requires only one correct alert during a lesion’s visible interval; the corresponding frame-level sensitivity of 52.8% shows that this recognition was not continuous. Model B was also trained and validated on the same processor platform within one institution; thus, some of its advantages may reflect domain proximity.

This false-positive trade-off deserves further scrutiny. Notably, a mean alert duration of 0.36 seconds does not by itself establish an absence of distraction: salience, timing, clustering, and the endoscopist’s response all matter, none of which are judgeable from an offline video. More fundamentally, a standalone benchmark measures the model and not the human–machine system. In a pragmatic randomized trial in community practice, CADe raised the SSL detection rate from 2.0% to 4.7%; this improvement is real, but remains far removed from what standalone sensitivity implies.8 Prospective evaluation should therefore report SSL detection rate, SSLs per colonoscopy, tandem miss rate, withdrawal time, and alert burden, alongside the human-factor outcomes that decision support systems driven by artificial intelligence (DECIDE-AI) asks for from early clinical AI studies.9

A broader lesson which should be considered is that private datasets are not superior to public datasets. Public repositories remain indispensable for reproducibility and independent scrutiny; carefully governed clinical datasets supply the pathological linkages and procedural diversity that public datasets lack. The REAL-Colon database, which includes multicenter, full-resolution video with expert annotation and histopathologic metadata, released for transparent benchmarking, shows that the two can coexist.10 Building consented, pathology-linked registries from high-detecting endoscopists is a task for societies and health systems, not individual vendors. The field would also benefit from declaring training-data provenance (label basis, contributing endoscopists’ detection rates, corpus composition) as routinely as the model architecture and from reporting the per-histology sensitivity, rather than folding SSLs into a generic “polyp” endpoint.

Han et al.1 offered a persuasive proof-of-concept that data curation deserves as much attention as model design. How much of their gain came from curation versus a corpus twice the size remains unresolved. Clinical superiority has not yet been established, but the direction is correct. For lesions as subtle as SSLs, the next advancement in CADe may come less from a deeper network than from a better-defined ground truth, richer clinical context, and validation built to reflect the complexity of real-world colonoscopy practice.