ABSTRACT
Pressure ulcers are the clinicopathology of local damage to the skin and underlying tissues caused by pressure, shear, or friction. Diagnosis, care, and treatment of pressure ulcers can be extremely expensive for healthcare systems. Reliable diagnosis and accurate assessment are critical to successful treatment decisions and, in some cases, to save the lives of patients. However, the current clinical evaluation procedures, which mainly focus on the visual inspection of the wounds, seem to be insufficiently accurate to accomplish this important task.
In recent years, the use of machine learning and deep learning techniques has aimed at automating wound diagnosis and providing a solution to this issue. Despite the large number of studies in the field of clinical image segmentation and the proliferation of deep learning models, their application to wound diagnosis is not always as direct or effective as expected. Since the number of wound image samples available in a regular dataset is usually very small, the direct application of deep learning-based segmentation techniques is limited.
In this work, we apply a strategy to counteract this issue by generating a significantly larger dataset, made up of the regions of interest (ROIs) extracted from the images and associated with the different areas of tissues present in the wounds. Thus for model training, hundreds of ROIs are generated from the ideal segmentation (ground truth) of each pressure ulcer image, where the label or class of the central pixel of each ROI is used to label the entire region. Later, a multiclass classification model based on convolutional networks is trained, where each region is classified into a class based on the severity of the area (tissue type). Finally, the final segmented image is reconstructed using the output of the neuronal model for each region. For model testing, overlapping regions are generated from each input image that is to be given to the model and are used to reconstruct the final segmented image. A small-size wound image dataset is utilized in this approach, consisting of 113 images of pressure ulcers in which five different tissue types can be found: skin, healing, granulation, slough, and necrosis. This dataset is made up of real clinical images obtained by healthcare professionals during their daily work on wound evaluation and diagnosis.
