WoundNet: A Domain-Adaptable Few-Shot Classification Framework for Wound Healing Assessment

Early detection of whether a wound is a "healer" or a "non-healer" using image analysis enables healthcare professionals to administer appropriate interventions. We propose a few-shot wound healing assessment framework, WoundNet, to classify temporal wound image sequences as "healer" or "non-healer." The contributions of this work are twofold: 1) Meta-learning: We study transfer learning approaches to train an image encoder for enhanced feature learning using domain adaption and contrastive learning, and 2) Few-shot classification: We classify image embeddings based on latent space similarity. We analyze the performance of the WoundNet framework in various settings, and experimental results show that temporal wound image sequences can be classified with an accuracy of up to 92%.