Deep neural network-driven infrared thermography for diagnostics and therapy monitoring of peripheral arterial disease

Infrared thermography (IRT) may offer a cost-effective, non-invasive method to support peripheral arterial disease (PAD) diagnosis and therapy monitoring. Although IRT is promising in PAD, single thermal images and certain features are often only analysed manually, which is time-consuming, hampers reproducibility, and the utilisation of the full potential of IRT in clinical practice. Therefore, we developed a deep neural network-driven processing pipeline (DNN) of specific cardiovascular-associated features (e.g., superficial veins & cutaneous arterial perforator vessels) within thermal images [1,2]. Consequently, the study aimed to test and explore the time series of automatically extracted thermal distribution patterns in patients with PAD during walking before and after interventional revascularisation.

In the present pilot study, three participants with PAD in Fontaine stage IIa – IIb (female aged 57 years, female aged 47 years, male aged 65 years) performed a treadmill walking test before and one week after interventional revascularisation. High-resolution IRT continuously measured the thermal distribution of the skin in °C in the posterior legs before, during, and after walking. Furthermore, we assessed the ankle-brachial pressure index (ABI), perceived exertion rate (RPE [scale 6-20]), and walking distance [m]. Our deep neural network-driven thermal image processing pipeline automatically analysed thermal images (~15.000 per test) and provided a time series of thermal features, including thermal distribution data for mean temperature (TS), superficial vein temperature pattern (TV), and cutaneous arterial perforator vessel temperature pattern (TP).

The measurement system successfully captured the decrease in TS, TV, and TP during walking. Revascularisation improved the ABI before (+0, +0.6, +0.2) and/or after walking (+0.5, +0.1, +0.3) as well as the RPE while achieving an equal or longer walking distance (-2, -5, -2). TS, TV, and TP decreased more pronouncedly during walking in two participants after revascularisation (TS: +0.38, +0.38, -0.05; TV: +0.39, +0.91, -0.34; TP: +0.58, +1.06, -0.05) (s. Figure 1.). Time series of all thermal distribution classes showed comparable inter-individual initial decrease but different variations during the course of the test (s. Figure 2.). In contrast, intra-individual relative thermal distribution variations were similar between the first and second walking tests.

DNN-IRT enables an automatic, reproducible and precise time series analysis of thermal distribution variations during walking, which could indirectly indicate improved tissue perfusion in patients with PAD. Extensive research is needed to generate large data sets to allow falsification of preliminary assumptions or identification of reference values and to develop classification algorithms to advance this new technique aiming for implementation in clinical practice to improve diagnostics and monitoring of PAD patients.Figure 1.

Region of interest