Segmentation Approaches for Diabetic Foot Disorders

Abstract

Thermography enables non-invasive, accessible, and easily repeated foot temperature measurements for diabetic patients, promoting early detection and regular monitoring protocols, that limit the incidence of disabling conditions associated with diabetic foot disorders. The establishment of this application into standard diabetic care protocols requires to overcome technical issues, particularly the foot sole segmentation. In this work we implemented and evaluated several segmentation approaches which include conventional and Deep Learning methods. Multimodal images, constituted by registered visual-light, infrared and depth images, were acquired for 37 healthy subjects. The segmentation methods explored were based on both visual-light as well as infrared images, and optimization was achieved using the spatial information provided by the depth images. Furthermore, a ground truth was established from the manual segmentation performed by two independent researchers. Overall, the performance level of all the implemented approaches was satisfactory. Although the best performance, in terms of spatial overlap, accuracy, and precision, was found for the Skin and U-Net approaches optimized by the spatial information. However, the robustness of the U-Net approach is preferred.

Keywords: diabetic foot (D017719); diabetic neuropathy (D003929); segmentation; supervised and unsupervised algorithms; thermography (D013817).

Figure 1

Simplified workflow for the U-Net + Depth (UPD) approach for IR image segmentation.

Figure 2

Illustrative example showing the input images and the final prediction including the improvement achieved by including the depth information: (a) RGB input image; (b) Depth input image; (c) IR original image and overlapping U-Net final prediction (white mask); (d) IR original image and overlapping UPD final prediction (white mask). The ground truth was overlapped (pink mask) to facilitate the comparison with the predictions.

Figure 3

Example showing the final prediction for the Skin approach in which the ground truth has been overlapped (pink mask): (a) RGB input image; (b) Depth input image; (c) IR original image and overlapping Skin final prediction (white mask); (d) IR original image and overlapping SPD final prediction (white mask).

Figure 4

Illustrative example showing the final prediction for the SegNet approach according to the $\alpha$ parameter in Equation (1): (a) IR input image; (b) Final prediction $\alpha$ = 0; (c) Final prediction $\alpha$ = 0.2; (d) Final prediction $\alpha$ = 0.4; (e) Final prediction $\alpha$ = 0.6; (f) Final prediction $\alpha$ = 0.8; (g) Final prediction $\alpha$ = 1. The ground truth (pink mask) was overlapped to both, the IR image and the corresponding final predictions (white mask).

Figure 5

Illustrative examples showing the final prediction for the SegNet approach ($\alpha$ = 1): (a) IR input image and the best prediction at T0; (b) IR input image and the best prediction at T5; (c) IR input image and worst prediction at T0; (d) IR input image and worst prediction at T5. The ground truth (pink mask), was overlapped to both, the IR and the final prediction (white mask).