Abstract
Accurate monitoring of chronic wound progression is crucial for assessing healing dynamics. However, the current manual process of tissue segmentation and quantification, which is an indicator of the healing progress, is time-consuming and subject to variability, so automated methods that can effectively monitor wound healing are required. In this work, inter-rater agreement analyses were conducted to evaluate the consistency of manual annotations performed by multiple experts and an automated methodology for tissue segmentation leveraging advanced deep learning techniques is proposed. For this, the convolutional neural network DeepLabV3-R50 and a transformer-based approach (SegFormer-B0) were explored. Furthermore, the potential of transferring knowledge from open wound segmentation models trained on different available datasets and fine-tuning them for this specific task was investigated. The tissue segmentation model is integrated into a framework that combines a wound and reference marker detection model with a wound segmentation model to refine the predicted tissue masks. The results show the benefits of employing previous knowledge gained from simpler tasks within the same domain, as well as the efficacy of post-processing operations. The top-performing tissue segmentation model, DeepLabV3-R50, achieved an overall mean Intersection over Union of 62.95% and a mean Dice score of 76.82% for the three analysed tissues on the Wounds dataset, when assessed independently. Considering the complete framework, the same model returns a mean Intersection over Union of 59.67% and a mean Dice of 74.38%, resulting in mean absolute errors of 14.33%, 14.31% and 8.84% for granulation, slough and eschar proportion estimation, respectively. Moreover, the obtained inter-rater agreement scores still emphasize the inherent complexity of the task, as even experienced healthcare professionals may differ in delineating tissue boundaries. Given the proven intricacy of tissue characterisation and the promising results that were achieved, the proposed pipeline contributes to streamline the tissue segmentation and quantification task, leading to the automation of the wound bed characterisation process and enhancing consistency and efficiency in wound healing.