JMIR Publications

ABSTRACT

Background:

Chronic ulcers are usually the result of prolonged pressure on the skin and underlying tissues. The assessment and treatment of wounds require an accurate analysis of their physical characteristics. In most cases, the methods of analysis used nowadays are rudimentary, which leads to errors and the use of invasive and uncomfortable techniques for patients. There are many studies that include a Convolutional Neural Network for classify the different tissues in a wound. In order to obtain good results the network has to be trained with a correctly labeled data set by a professional expert in wound assessment. Typically, it is a hard work to labelling pixel by pixel using a professional photo editor software. It needs a long time and effort by health professional.

Objective:

The main goal is implementing a new fast and accurate method of labelling wound samples for training a neural network to classify different tissues.

Methods:

Developing a support tool and evaluate the accuracy and reliability. Compare the support tool classification with a digital gold standard.

Results:

To evaluate this new method, it has been taken as reference the gold standard for labeling image segmentation data. In that case, labeling the data with an image edition software. The obtained average ratio of each tissue type are 0.3759 for no skin, 0. 5935 for healty skin, 0. 0178 for granular and 0.0110 and 0.0001 for base and necrotic.

Conclusions:

This method increases tagging speed an average compared to an advanced image editing user. This gain is bigger with untrained users. The samples obtained with the new system are indistinguishable from the samples made with the gold standard.