Automated Segmentation of Wound Bed Tissue from Clinical Photographs: A Benchmark of Five Deep Learning Architectures Under a Shared Protocol

Accurate characterisation of the wound bed, the relative proportions of granulation, slough and necrotic tissue described by the red–yellow–black model, is central to monitoring healing and guiding debridement in chronic-wound care, yet in routine practice, it is judged visually and is poorly reproducible between clinicians. We investigated automatic, pixel-level segmentation of these three classes within the wound region of interest. A corpus of 404 photographs was curated from a public wound image collection and densely annotated by wound care clinicians with a purpose-built, cluster-assisted labelling tool. Five deep learning architectures spanning convolutional, transformer, hybrid and ensemble families were trained on one frozen data split under a largely shared pipeline, with limited, fully documented architecture-specific tuning, and scored on a shared 62-image test set. The models fell within a narrow mean intersection-over-union band (0.748–0.818). The strongest single model, a purely convolutional network, was statistically indistinguishable from the best ensemble, and simply averaging already-trained convolutional and transformer models matched that ensemble at no additional training cost. Slough remained consistently the hardest tissue. Translated into clinical units, per-wound tissue percentages were recovered with small mean bias (within ± 5 percentage points) but wide single-image limits of agreement (about ± 25–50 percentage points), so the measurement suits aggregate and longitudinal assessment rather than single-image reading. Within this dataset, split and training protocol, the choice among the strong hierarchical architectures was not decisive: a single strong model is the pragmatic option for clinical deployment, with an ensemble reserved for maximal accuracy and class balance.