Automatic detection of risk zones in diabetic foot soles by processing thermographic images taken in an uncontrolled environment

Diabetes mellitus has become a global healthcare issue with incidence levels growing exponentially each year. Diabetic foot is one of the complications related with the ailment, and if it is not attended in time, it can progressively deteriorate to a condition that necessitates foot amputation. This study aimed to establish a non-invasive monitoring system for diabetic foot. This proposed system detects and classifies temperature differences in foot sole zones as ulcerous if >2.2 °C and necrotic if <-2.2 °C, by processing thermographic images of foot soles. This is achieved without homogeneous background or room temperature control. There are reports of systems that are designed to work under controlled environments; however, their performance declines or falters altogether under uncontrolled, home environments.

The system proposed in this paper combines step-by-step and end-to-end algorithms that compensate for the limitations in data and, at the same time, enhance performance. It uses deep-learning techniques to segment visible-spectrum images using a retrained Mask R-CNN model, which is adjusted with 141 images to segment foot soles. These results are used over the temperature matrix in order to isolate the foot sole temperatures. The visualisation and classification methods use step-by-step algorithms with comparisons of homologous foot sole regions, convolutions with a two-dimensional (2D) Gaussian function, filters that process and compare areas of 1.5 cm² over each foot sole, and a thermal threshold to differentiate ulcerous from necrotic zones.

The results illustrated a detection accuracy of 90% for ulcers and 88% for necrosis, while the labelled areas had an error of 7.05% and 10%, respectively. These results demonstrated that the system is capable of successfully detecting and visualising the specific temperature differences over samples under an uncontrolled environment.