Evaluation of Peripheral Arterial Disease Through Thermal Imaging and Machine Learning: A Texture- and Frequency-Based Approach

CARLAK, HAMZA; ÖZÇOBANOĞLU, SALİH; Irim, Yakup

doi:10.1007/s13369-026-11290-2

Evaluation of Peripheral Arterial Disease Through Thermal Imaging and Machine Learning: A Texture- and Frequency-Based Approach

CARLAK H. F., ÖZÇOBANOĞLU S., Irim Y.

Arabian Journal for Science and Engineering, 2026 (SCI-Expanded, Scopus)

Yayın Türü: Makale / Tam Makale
Basım Tarihi: 2026
Doi Numarası: 10.1007/s13369-026-11290-2
Dergi Adı: Arabian Journal for Science and Engineering
Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, zbMATH
Anahtar Kelimeler: Convolutional neural network, Image processing, Peripheral arterial disease, Thermal Imaging, XGB model
Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
Akdeniz Üniversitesi Adresli: Evet

Özet

Peripheral arterial disease (PAD) is a common cardiovascular condition. The aim of this study is to facilitate the early diagnosis of PAD in the lower limbs and to develop a non-invasive assistant medical diagnostic method to monitor the treatment process. The proposed method was developed using a combination of thermal imaging, image processing, and artificial intelligence algorithms. The proposed framework was evaluated on a cohort of 40 participants divided into four groups: athlete, healthy, individuals at cardiovascular risk, and patients with PAD. Infrared thermal images of the lower extremities were acquired under standardized conditions before and after a controlled physical effort protocol. Haralick texture features, including energy, contrast, correlation, homogeneity, and entropy were calculated. Hot region distribution images were generated using K-means clustering. Frequency components value of any signal was calculated by wavelet transform. The highest temperature increase due to effort was observed in the athlete group (ΔT = 0.96 °C), while the lowest was observed in the patient group (ΔT = 0.38 °C). The Kruskal–Wallis H test results showed statistically significant differences in contrast (p = 0.017) and homogeneity (p = 0.028) between groups. The Haralick texture features were trained using the Support Vector Machine algorithm, achieving a 95.7% success rate. Hot region images were trained using a convolutional neural network model, which also achieved a 95.7% success rate. Frequency features were trained using the K-Nearest Neighbors algorithm, achieving an 83.3% success rate. Subsequently, an eXtreme Gradient Boosting (XGB) model achieved a success rate of 92.9%. The technology explored in this study holds potential as an assistant method for detecting, monitoring, and evaluating responses to treatments for conditions such as peripheral atherosclerotic diseases, diabetic vascular diseases, and others.