Incremental value of a cross-dataset ulcer-region prior for infection and ischaemia classification in diabetic foot ulcer image patches

Objective To assess whether a DFUC2022-derived soft ulcer-region prior improves four-class infection/ischaemia label prediction in ulcer-centred DFUC2021 patches, using comparators matched for classifier architecture or base-model outputs. Methods We analysed 5, 955 labelled patches using five fixed image-similarity-grouped outer folds and five training seeds. CNN checkpoint selection and stacker fitting and selection excluded each outer evaluation fold. Three incremental comparisons assessed Full lesion-prior transfer (LPT) against an architecture-matched neutral-prior control, reliability-aware LPT (RA-LPT) against its output-only stack, and strong RGB stacks with versus without explicit prior features. Pointwise 95% confidence intervals and approximate two-sided p values were obtained from 5, 000 paired crossed seed × similarity-group bootstrap resamples; p values were Holm-adjusted across the three comparisons. Results Full LPT achieved Macro-F1 of 0.8921 ± 0.0059 versus 0.8777 ± 0.0085 for neutral Full LPT (mean ± sample SD across seeds). The mean paired difference was +0.0144 (95% CI, + 0.0048 to +0.0244; Holm-adjusted approximate p = 0.0132), positive in all five seeds. RA-LPT achieved Macro-F1 of 0.9181 ± 0.0038 versus 0.9206 ± 0.0028 for its corresponding output-only stack, without demonstrating an incremental improvement (difference, −0.0026; 95% CI, −0.0066 to +0.0012; adjusted p = 0.3079). Adding explicit prior features to the strong RGB stack also did not demonstrate an improvement in Macro-F1 (−0.0016; 95% CI, −0.0057 to +0.0025; adjusted p = 0.4223). The strong RGB stack achieved 0.9370 ± 0.0050, above Full LPT. In a later exploratory training-size analysis, Full LPT and its neutral control improved with increasing labelled-training budgets; mean Full LPT-minus-neutral Macro-F1 differences were positive at all four budgets, but all pointwise 95% confidence intervals included zero. Conclusion In the main five-seed evaluation, sample-specific prior information improved Full LPT relative to its architecture-matched neutral control. Its incremental value was comparator-dependent: additional explicit prior features showed no confirmed improvement beyond the evaluated model-output stacks. Patient identifiers and target-domain lesion masks were unavailable; these internal development findings require patient-linked, multicentre external validation before clinical use.