Actinic keratosis labeling and single-band instability in near-infrared spectroscopy of skin cancer: a nested, patient-level re-analysis

Labeling actinic keratosis (ACK) in the NIR-SC-UFES skin-lesion dataset changes macro-F1 from 0.764 (ACK non-cancer) to 0.677 (ACK positive), more than classifier choice.

• Why it matters: Benchmark interpretation is unclear because ACK labeling conventions and the stability of ranked spectral bands may bias reported performance and hinder spectroscopic explanations.
• What they did: Using 714 in vivo spectra (331 patients; 908–1676 nm) with SNV normalization, five classifiers were assessed via nested patient-level cross-validation, while band importance was tested with 60 patient-level bootstrap resamples and band-removal experiments.
• The result: Top-10 band importance was unstable (Jaccard 0.34; Spearman 0.51), though the top-ranked band lay in 908–1051 nm in 59/60 resamples; removing top 5–20 bands reduced macro-F1 by at most 0.016.