I. Introduction
Diabetic foot is a common disease among diabetic patients leading to foot ulceration which is the primary reason for diabetic related foot amputations. The occurrence of foot ulcer is often associated with foot hyperthermia. In [1], foot hyperthermia was defined as a temperature difference higher than 2.2°C between a foot region and the same region on the contralateral foot. According to [1], foot ulcer occurrence can be reduced by 70% if the foot hyperthermia is early detected. This information can be identified using a thermal infrared camera, as reported in several studies [2]–[5]. In fact, most of these studies require a constraining acquisition protocol which imposes on the person who participates in the acquisition to put his feet in a special device that hides all other thermal sources except those coming from the plantar foot. This process ensures good image quality and therefore a low level of segmentation complexity. Our aim is to develop new mobile and user-friendly technology to precisely analyze the plantar foot temperature in diabetic foot problem. We, therefore, free ourselves from using a complex and constraining isolation system unlike [3]–[6]. Images will be freehandly taken with a smartphone equipped by a dedicated thermal camera. The full automatic processing of the data is to be performed in the smartphone itself. It is expected that this overall protocol can be generalized in a clinical routine or even at home. The first step of this processing is a fully automatic segmentation of the plantar foot surface. Thus, the automatic segmentation of such images is difficult, since the occurrence of other thermal sources of the body. Classical segmentation methods fail to segment these specific images as demonstrated in a recent work [7]. We, therefore, presented a prior shape based active contour method that proposes to modify the snake functional of Kass et al. [8] by adding an extra energy term. This term guides the snake to the desired contour by minimizing a curvature difference between the snake curve and the prior shape curve. In previous work [9], we compared the proposed method to two prior shape-based active contour models. The first method is the Ahmed et al. method [10] which assesses the shape matching performed directly in the Fourier descriptor space. The second method is proposed by Chen et al. [11]; authors proposed to find the transformations (scale, rotation, and translation) such as the prior shape is closely associated with the transformed curve. When applied to our database of 50 plantar foot thermal images, results show that our proposed method outperforms the two others. The major problem with these methods is their sensitivity to the position of the initial contour. Even though the previosly proposed method is less sensitive, an automatic initialization process has to be performed to obtain a better result. Moreover, the characteristics of the images are different from person to person. Let’s take the example of a healthy (or not) person with toes that are always cold. In this case, active contourbased methods may fail to find the good foot contour even with an imposed shape constraint. So that, more powerful and robust segmentation methods are required. In recent years, deep learning techniques have proven spectacular progress. Initially intended for image classification [12]–[15], they are more and more applied to a wide variety of other tasks, in particular for semantic segmentation. To the best of our knowledge, no studies have applied deep learning for diabetic plantar foot thermal images segmentation.