Thermographic images are widely used in diagnosing Diabetic foot (DF) infection and ulceration that are associated with neurological abnormalities. Thermographic images of various diabetic foot ulcerated conditions are collected from 60 patients. There are twenty different types of conceptual foot patterns occurring in diabetic patients, out of which most predominantly observed are type 3A, type 1D, type 3D, type 4D. This work proposes on developing an automatic algorithm to match these diabetic thermographic images of ulcerated foot with conceptual template rather than manually comparing it with the thermogram datasheet. A conceptual template is created using the thermogram datasheet for each of these twenty patterns by edge mapping and curvelet technique. About 60 thermographic images of diabetic foot that have the most predominant patterns are taken for study. These images are pre-processed and segmented for ulcerated region of interest. A pattern for these thermographic images is generated using edge mapping technique and Hough transform. The broken edges are joined by curvelet technique to form a pattern. Each of these patterns is compared with the twenty conceptual templates by template matching technique and an overall efficiency of about 95.27 percent is achieved. This automated algorithm helps the physicians in finding the accurate type of diabetic foot and it is possible to detect the pre-signs of ulcerations in foot.

Diabetes mellitus is one of the most commonly diagnosed disorders around the globe. Around 400 million people globally are affected with this disorder with around 1.2 million deaths. Some complications include diabetic retinopathy, foot ulcers, and renal failure. The use of thermal imaging makes the diagnostic process non-invasive, reducing discomfort and the risk of infection associated with traditional methods. Early detection of diabetic foot ulcers is crucial for timely intervention and prevention of severe complications, such as infections and amputations. Deep learning models have the potential to provide accurate diagnoses by analyzing thermal patterns that may not be easily discernible by the human eye and could potentially track the progression and severity of the disease over time. A total of 996 images with 264 healthy images and 732 unhealthy RGB images were used in this work. The total dataset was split into training, testing, and validation groups in the ratio 6:2:2. The left and right foot images were equally randomized in the respective models while training. ResNet50 is used with a learning rate of 0.01 having 20 epoch gives the highest accuracy of 98% and a minimum loss of 1% further, the work can be enhanced with a real-time database.