3D image analysis of foot wounds

Foot wounds are a debilitating and potentially fatal consequence of diabetes. Assessment of foot wounds in clinical or research settings is often based on subjective human judgement which does not involve quantitative measurement. When measurement is conducted, it takes the form of ruler-based estimations of length and width to approximate perimeter or area. To monitor wound healing and make informed treatment decisions, clinicians require accurate and appropriate measurements of wound parameters. Effective wound assessment requires imaging and software techniques which enable objective identification of wound tissues and three-dimensional measurements of wound size. Pilot classification studies were carried out using a selection of six stock wound images. Ground truth was provided by a specialist practitioner in podiatry. Three supervised classifiers were compared. Maximum likelihood was found to be the most suitable for wound classification. Performance of the supervised Maximum likelihood (MLC), unsupervised Expectation Maximisation (EM) algorithm and a hybrid MLC-EM method were compared. No method was found to perform significantly better than others. Context classification was implemented via probabilistic relaxation labelling. It was found that classification accuracy was typically improved by 0.5 - 1.5 %. A method of including depth information in the classification process was proposed and evaluated. Simulated 3D wound volumes were imaged and combined with simulated tissue colours sampled from real images. Classification using depth improved accuracy at low weightings when included in the Maximum likelihood classifier. To facilitate the further development and evaluation of novel wound assessment algorithms, a set of clinical foot wound data was imaged using 3D stereophotogrammetry. A group of clinicians assessed the data to identify the tissues contained within each wound image. The level of agreement between them was evaluated. Supervised, unsupervised and hybrid classification algorithms were also used to classify the data and the results were evaluated by comparison to the group of clinicians. Novel methods of measuring the volume and surface area of wounds were developed and validated using simulated models before being applied to wound data. The results of tissue classification were plotted against the results of volume measurement in order to observe any trends in the healing process. Supervised Maximum Likelihood classification was found to produce results which agreed with clinicians to approximately the same level as they agreed with each other, indicating that automated classification may have a future role in wound research and clinical diagnosis. The supervised method resulted in agreement with clinicians of 75.5%, which was significantly higher than agreement for unsupervised or hybrid methods, at 65.9% and 64.6% respectively. The inclusion of tissue depth in the classification progress produced some positive results. The surface area and volume measurement methods were found to be accurate for all but the smallest of wound sizes and capable of tracking changes in real wounds.