PFUTnet: A Novel Deep Learning Architecture for Diabetic Foot Severity Mapping and Analysis

• Published in: IEEE sensors journal Vol. 24; no. 9; pp. 14770 - 14777
• Main Authors: Sharma, Naveen, Mirza, Sarfaraj, Rastogi, Ashu, Mahapatra, Prasant Kumar, Kumar, Deepak
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.05.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Asymmetric analysis; classification; Classification algorithms; Deep learning; deep learning (DL); Diabetes; diabetic foot; Diabetic neuropathy; Feature extraction; Foot; Foot diseases; Image segmentation; Machine learning; Males; plantar foot ulcer thermogram network (PFUTnet); Sensors; Temperature distribution; thermal image; Ulcers
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2024.3377212

Clinicians traditionally rely on manual inspection to assess the severity of diabetic foot ulcers (DFUs). However, this approach is subjective and prone to errors. Effective deep learning (DL) classifiers may minimize the users' dependency. In this manuscript, we propose a randomized, prospective, single-blind study focusing on diabetic neuropathy foot ulceration. The study included 104 subjects, with 71 patients (50 males) having DFU (Wagner grade 2) on the plantar aspect of the foot, while the remaining 33 subjects were nondiabetic (controlled) individuals (20 males). The average age of the patients was 54.28 ± 7.45 years, and the duration of the ulcers was 5.86 ± 2.22 months. The plantar foot was divided into three regions (forefoot, midfoot, and hindfoot) to detect ulcers in various areas. A Novel DL model, plantar foot ulcer thermogram network (PFUTnet), was introduced to evaluate the severity of each area. When compared to existing DL models, such as Inception V3 and AlexNet, PFUTnet demonstrated superior performance with a higher area under curve (AUC) score of 0.98 on the thermo dataset. In terms of classification accuracy, PFUTnet achieved approximately 95%, while Inception V3 and AlexNet achieved around 93% and 90%, respectively. The proposed technique enables automated and user-independent assessment of diabetic foot wounds, allowing for precise characterization of the severity of diabetic foot conditions.