Machine learning approaches for diabetic foot wound segmentation: generative models vs. CNNs vs. K-Means

Current medical applications of automatic image segmentation include detecting and subtracting meaningful information from regions of interest to estimate physical and physiological parameters related to the human body. However, clinical tracing is still required to validate its performance. This research hypothesized that a new approach known as Attention Diffusion Zero-shot Unsupervised System (ADZUS) can be used for the segmentation of wounds without the use of labeled data as reference. Herein, a comparison between the proposed approach against Otsu’s method, two unsupervised machine learning algorithms for clusterization, K-Means, and density-based spatial clustering of applications with noise (DBSCAN), and convolutional neural networks (CNN) is presented to evaluate the segmentation performance. A set of diabetic foot ulcer images, acquired by following a clinical protocol, was used to test the segmentation performance of each technique. Intersection over Union (IoU) and Dice Similarity Coefficient (DSC) were used to compare the performance of each algorithm based on the segmented image against labeled clinical tracings. The analysis showed that ADZUS performance (IoU=0.51, DSC=0.65) outperformed K-Means (IoU=0.44, DSC=0.62) and CNN (IoU=0.47, DSC=0.64). The proposed wound segmentation approach using ADZUS is innovative since it doesn’t rely on the availability of labeled images with clinical tracings or the inclusion of large datasets of clinical images. This advantage overcomes the use of conventional machine learning techniques commonly used as segmentation tools. Ongoing studies are the implementation of ADZUS to segment different tissue types and to apply skin color correction to periwound regions for imaging studies of diabetic foot ulcers.