Inside AI Tissue Classification: How Models Separate Necrotic, Granulation, and…

Tissue classification is the workhorse of wound care AI. Getting the percentage of necrotic, slough, granulation, and epithelializing tissue right drives debridement decisions, healing scoring, and reimbursement documentation. Here is how the models actually work.

The input: a standardized image

Classification only works on a clean input. Point-and-capture workflows enforce consistent distance and lighting, and AI-powered wound imaging writes fiducial-corrected pixels to the model. Camera-roll snapshots taken at random angles are the largest source of downstream error.

The architecture: semantic segmentation

Modern classifiers use encoder-decoder convolutional networks (U-Net variants) or vision transformers that assign every pixel to a tissue class. The output is not a single label but a per-pixel map — from which percentage composition, hotspot location, and change-over-time are computed.

The training data problem

Accuracy above ninety percent is reported on validated cohorts, but generalization depends on training diversity: skin tone, wound etiology (DFU, VLU, pressure, surgical), anatomic site, and lighting conditions. Models validated only on light skin or only on DFUs will fail silently on the rest.

Where models still struggle

• Biofilm — visually subtle, often misclassified as granulation.
• Eschar vs deep necrosis — depth cues are limited from 2D images.
• Maceration boundaries — periwound classification remains noisier than wound-bed.

From classification to clinical decision

A pixel map is data, not care. Feeding tissue composition into a healing dashboard turns it into trajectory: rising granulation with falling slough predicts closure; stable necrosis predicts stall. That is where wound healing software earns its keep.

Tissue classification is mature enough to trust as a measurement aid — with a validated model, standardized capture, and a clinician reading the output in context.