Deep neural net for identification of neuropathic foot in subjects with type 2 diabetes mellitus using plantar foot thermographic images

• Published in: Biomedical signal processing and control Vol. 96; p. 106509
• Main Authors: Christy Evangeline, N., Srinivasan, S.
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.10.2024
• Subjects: Convolutional neural nets; Custom deep net; Diabetic foot; Diabetic neuropathic foot thermal images; Custom deep net; Diabetic neuropathic foot thermal images; Convolutional neural nets; Diabetic foot
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2024.106509

•Diabetic neuropathy is a serious stimulus to diabetic foot ulcers which in turn may end up in infections and amputations in the lower limb.•This work has harnessed the edge of asymmetry analysis in identifying neuropathic feet.•The work confers efficacy of deep learning in clinical decision support.•The study compares a custom deep learning net with off-the-shelf nets with respect to this specific application.•Seemingly, this work is unprecedented by any attempts to classify healthy and neuropathic feet using thermal imaging modality. Diabetic foot syndrome (DFS) is characterized by anomalies in thermal distribution in plantar foot of diabetic subjects. While non-diabetic subjects exhibit symmetric thermal patterns between both feet, DFS subjects demonstrate deviation from symmetricity. These temperature patterns, can be captured through infrared camera that can be supportive in identifying subjects with diabetic peripheral neuropathy (DPN). DPN is one of the main factors that engenders diabetic ulcers that in most cases end up in infections and amputations. This makes timely detection of DPN, a precondition in diabetic foot care. This demand necessitates the application of infrared camera in classification of DPN subjects from normals. This paper compares the suitability of off-the-shelf nets – MobileNet and ResNet-50 in identification of DPN. Also, a custom convolutional neural net (CNN)-based deep network, DPN-Net, was developed to perform classification and was compared with other nets basis their performance. DPN-Net achieved 98.5 % test and 98.7 % validation accuracies with comparatively lesser trainable parameters making DPN-Net suitable for a telemedicine mobile application. Presumably, this work is the first in place to perform classification between healthy and neuropathic feet using thermography. The work aims to ascertain foot health in diabetic subjects and conjointly presents comparison between off-the-shelf and the custom net in carrying out the classification. The performance of DPN-Net proved better than off-the-shelf nets showing that the classification can be achieved over a comparatively simpler net, thereby reducing computational cost and complexity.