DETECTION AND CLASSIFICATION OF DIABETIC FOOT THERMOGRAMS USING DEEP LEARNING

Abstract

Diabetes is a long-lasting disorder which is also a cause for high mortality rate. Diabetic Foot Ulcers (DFU) is a condition associated with diabetes. Diabetes, a chronic ailment generating high blood sugar is a key reason for mortality and lifestyle degradation. Thermography is an emerging and non-invasive medical imaging practice that can be used for body disorder diagnosis. Temperature difference from the usual body temperature is the basis of this technique. In the recent scenario, Deep Learning (DL) algorithms serve as a smart and precise approach for classifying DFUs with the utilization of thermograms. In this work, performance of transfer learning based DL networks (AlexNet and Resnet-101) has been analyzed in classifying foot thermograms into CG (Control Group) and DM (Diabetic Mellitus) groups. Furthermore, a novel and enhanced technique (ProNet), is proposed in this paper. It is designed by combining the best features of both AlexNet and ResNet. By simulation results obtained in this work, it can be inferred that AlexNet has greater accuracy (96.8%) than that obtained in the existing work. In addition to this, ResNet101 yields better results than AlexNet (accuracy of 97.9%). ProNet, the proposed methodology achieves improved performance in terms of accuracy (as high as 98.9%) and other metrics like precision, specificity and F1 Score. Also, this work evaluates the performance of transfer learning DL networks (DarkNet-19 and DarkNet-53) in multiclass classification of DFUs based on Thermal Change Indices (TCIs). Moreover, an improved and novel methodology (Pro-Multi-Net), is proposed. The optimum features of both DarkNet-19 and DarkNet-53 are combined to design this algorithm. As per simulation results, it can be deduced that all the three methods implemented have higher accuracy than the techniques applied in the existing work. Furthermore, DarkNet-53 yields better results than DarkNet-19 (overall accuracy of 92.8%). Pro-Multi-Net, the proposed network achieves an improved overall accuracy of 95.5%. Also, precision, recall and F1-Score values are also obtained for all the classes. vi • Chapter-1 gives an introduction to diabetic foot ulcers, infrared thermography and anomaly detection procedure in diabetic foot. Moreover, it provides a brief outline of DFU classification using multiple DL techniques. • Chapter-2 discusses the literature review of work done in this domain along with a glimpse of proposed work is discussed in this chapter. • Chapter-3 explains the techniques implemented in this work for binary classification and multiclass classification of diabetic foot thermograms including the novel and proposed methodology. • Chapter-4 includes the results and discussion. It provides the simulation details and explains the results obtained in this work. Also, a comparative analysis of all the methods implemented is given. • Chapter-5 explains the conclusion and future work that can be carried out in this domain.