Open Access
Add to BibSonomy
Abstract
Diabetic foot is a well-known problem among patients suffering from peripheral arterial diseases (PAD). This article presents an optical sensor for contactless measurement of the anatomical site based on laser speckle techniques. The sensor illuminates the inspected tissue and analyzes the captured back-reflected light from the time-changing speckle patterns. An occlusion test was implemented to provide a statistical parameter to differentiate between a low perfused and a healthy foot. A clinical study of 15 subjects was conducted. The video was analyzed by two methods: dynamic laser speckle (DLS) and laser speckle contrast analysis (LASCA). Data analysis included several classification models, where the KNN model exhibited maximum performance. These findings suggest that a simple and inexpensive system for PAD monitoring can be designed for home use and/or in community clinics.
© 2022 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Blood microcirculation is critical to controlling patients’ hemodynamic and thermal state, as well as for regulating metabolism. The physiological state can vary over time and is considered a reflection of individual health. For this reason, blood microcirculation is often characterized as a biomarker of tissue perfusion in the presence of vascular injury or disease. Clinicians need to be able to monitor the state of flow and its distribution in the compartments of the microvasculature in a fast and convenient manner and with minimal patient discomfort [1,2]. Even though microcirculation accounts for approximately 5% of the total blood volume, it plays an important role in several vital functions including the regulation of blood pressure, metabolic exchanges, homeostasis of interstitial fluids and thermoregulation. However, its key function is to ensure proper tissue nutrition. Thus, early detection of diabetic foot vascular calcification is crucial since it reveals the distribution of blood between nutritional and non-nutritional skin micro vessels [3,4].
Peripheral arterial disease (PAD) is a well-known disease [5]. Patients with PAD often suffer from decreased mobility, multiple morbidities, and low quality of life. The first symptomatic stage of PAD is usually manifested in intermittent claudication, which progresses to chronic limb ischemia (CLI) in 15% of these patients. PAD does not progress linearly from claudication to CLI in every patient. Sometimes patients develop CLI without ever having claudication. Thus, accurate and easily accessible diagnostic techniques are necessary for the early detection and treatment of PAD. There are several non-optic methods currently available for PAD diagnosis, most notably Vascular Ultrasound (VUS) [6], Magnetic Resonance Imaging (MRI) [7], Ankle Brachial Index (ABI) [8] and Transcutaneous oxygen tension (TcPO2) [9,10]. VAS, ABI and TcPO2 are contact based techniques, whereas MRI is a very expensive device and constitutes a costly procedure for HMOs . Other optical imaging methods for early detection include Near Infrared Spectroscopy (NIRS) [11], Laser Doppler Flowmetry (LDF) [3] and Laser Speckle Imaging (LSI) [12]. LDF is the most common non-invasive remote method for evaluating wound depth in PAD injuries and can be used to monitor skin blood flow. Another method that is attracting increased interest is Multispectral Imaging [13]. Its ability to detect light across the spectrum and spatial domains at a very high chromatic resolution makes this approach a clear candidate for diabetic foot ulcer analysis. However, these methods fail to capture changes in the micro-circulation over the course of wound recovery. LSI, also known as Laser Speckle Contrast Analysis (LASCA), is based on the analysis of speckle pattern images created by scattered light. The dynamics derive from the interactions between coherent photons and tissue that create fluctuating speckle patterns, which result in spatial blurring (contrast reduction) when averaged over a constant exposure time [14–16]. TcPO2 is an established method for assessing cutaneous perfusion that measures the transfer of oxygen molecules to the skin surface and has been applied to the detection of PAD in diabetic patients. Since TcPO2 assesses the area under the probe, multiple readings are sometimes used to provide richer clinical information beyond a single measurement. In addition, a regional perfusion index can be applied by dividing the foot TcPO2 reading by a baseline TcPO2 value measured from the patient’s chest, which makes it more reliable than a single TcPO2 reading [17,18].
This paper presents a clinical study assessing perfusion in the lower limbs of PAD patients. Two laser speckle-based methods – Dynamic Laser Speckle [19,20] and Laser Speckle Contrast Analysis (LASCA) [21,22] – were evaluated. The sections below describe the clinical setup and protocol, the measurement system optimization analysis and an in-depth comparison of these methods. Relative to existing methods, we propose a remote, contamination free device. We show how it can be achieved with a low-cost camera and an off-the-shelf laser.
The innovation is this work lies in proving the ability to apply laser speckle-based method for measurement PAD patients. We propose an occlusion protocol which can analyze the dynamic of blood flow once obstruction is released. Based on this transient state, we build a classification model to accurately differentiate between healthy and diseased limbs. Further, we demonstrate how system parameters can be optimized to reduce cost but still maintain high performance. This prototype system can thus lay the groundwork for developing an inexpensive apparatus for home and/or in community clinics.
2. Methods
The sample was composed of 15 patients (for a total of 30 limbs). Due to technical issues related to the development of this prototype, 7 limb measurements had to be discarded. The remaining 23 limb characteristics are summarized in Table 1. The limbs in the healthy group were from 6 males and 3 females (2 smokers). The limbs in the sick group were from 3 males and 1 female (1 smoker). The ABI index was only measured for Diabetic patients. In the Diabetic group, 3 patients had Diabetic Foot Disease (Neuropathic), and 1 patient had a Diabetic ulcer (vascular).
Table 1. Limb Properties
An occlusion test was conducted to distinguish between PAD patients suffering from low perfusion in their feet and a healthy control group. Each patient was assessed on the following protocol: 3 minutes of baseline, 3 minutes of lower limb occlusion (via a dedicated sphygmomanometer), and 9 minutes of post-occlusion time. The sphygmomanometer cuff was placed directly above the ankle on each limb. Occlusion was achieved by inflating the cuff ∼50 [mmHg] above the patient’s systolic blood pressure. Blood pressure was measured for each patient prior to the occlusion test with a standard Omron M500 blood pressure monitor. Cuff inflation was done manually and took ∼ 5 seconds to reach maximum pressure. Cuff deflation was done by rotating the air-release knob on the sphygmomanometer, which took ∼2 seconds to release all the pressure. To avoid any undesired movement, the patient’s limbs were immobilized on the bed with a dedicated pillow. Patients were asked to remain as still as possible throughout the entire duration of the measurement.
Flow analysis can be carried out with a high-speed video camera to calculate the second order autocorrelation function (ACF), ${g^2}(\tau )$, at each point in the image (each pixel on the image sensor). This method is dubbed the Dynamic Laser Speckle here, and its basic steps are visualized in Fig. 1, and described in detail in Golberg et al. [19]. The entire video was divided into time slices. In this study, each time slice was 32 × 32 pixels, and had a duration of 20 [ms]. Figure 1 (a) – (d) show how the decay time was extracted from typical time slice.
Fig. 1. (a) One ‘time slice’ from the reflected video pattern; (b) Typical intensity variation of a single voxel during one ‘time slice’; (c) Single voxel ACF ; (d) Mean ACF when averaging over region of interest (ROI) of pixels
The same video can also be analyzed in a slightly different way, known in the literature as LASCA, where the camera integration time is compared to the speckle decorrelation time, thus creating a blurring effect of the recorded image. The level of blurring is quantified by the speckle contrast, C, which is usually defined as the ratio of the standard deviation $\mathrm{\sigma }$ of the intensity I to the mean intensity, $\langle \textrm{I} \rangle$, of the speckle pattern (Eq. (1).) If there is little or no movement in the object, there will be only a little or no blurring [23].
The optical system also needs to be optimized to yield the proper speckle-to-pixel ratio. When an imaging condition is generated, a subjective speckle is formed. In this case the speckle size on the image plane of the camera is determined by Eq. (3) [24]:
where λ is the wavelength of light, M is the magnification of the imaging system, and f# is the system’s f number. Averaging the speckles within a single pixel area will reduce the contrast and hence reduce the Signal to Noise Ratio (SNR) of the ACF, as well as the SNR in the τ values. To satisfy the Nyquist sampling criterion and maximize the contrast, the speckle size was set to twice the pixel size.3. Results and discussion
The analysis depicted in Fig. 1 serves to calculate the decay time for the full recorded video of the occlusion experiment. For each temporal slice of 20 [msec] we calculated the decay time, designated here by τ, which was defined as the time it takes the curve to drop to 50% of its full range (from max to min values). Figure 3 depicts the typical measurement of 1/τ for the occlusion test and shows several graph parameters. Since the ACF decay rate is closely related to blood flow rate, 1/τ is proportional to flow itself. As shown in the occlusion region (minutes 3-6), there was no or very little flow; thus 1/τ decreased significantly relative to its baseline flow prior to occlusion.
Fig. 3. 1/τ results of the occlusion test for PAD and healthy controls AUC – Area Under the Curve (donated by dotted grey lines), RF – reference flow, Max – maximal point of flow overshoot, TR – time from occlusion release until the flow returns to the RF level, TM – time from occlusion release until the flow reaches its maximum level, TH – time from occlusion release until the flow reaches half its maximum level
We tested all the parameters shown on the graph: TR, TM, TH, MAX, (MAX + RF)/2, RF and Area Under the Curve (AUC) to differentiate the healthy limbs (control group) from the sick limbs (PAD patients). Classification models based on each parameter showed that the best estimator of the differentiation was the AUC. Once the cuff is released there is a “burst” of blood flow to the foot, which results in the over-shoot relative to the reference flow (RF) level prior to occlusion. Then, the blood flow returns to its steady state. Foot vascular calcification prevents the rapid return of the blood in the overshoot, post occlusive time, thus making the AUC the best estimator for distinguishing between healthy and diabetic feet.
After building 1/τ plots for all the limbs in the clinical trial and extracting the AUC for each plot, we used the MATLAB Classification Learner App to evaluate several classifiers to differentiate between healthy and diabetic feet. We examined four classification models: Ensemble (preset: Boosted Trees, method: AdaBoost, type: Decision tree), SVM (preset: Cubic, kernel function: Cubic, Kernel scale: Automatic), Tree (preset: fine, split criterion: Gini’s diversity index) and KNN (preset: fine, k value: 4, distance metric: Euclidean). Figure 4 summarizes the accuracy results.
The K-Nearest Neighbors (KNN) classifier was chosen for this analysis. We performed a 5-fold cross validation to avoid overfitting. A two-sample t-test was calculated to reject the null hypothesis, and to verify that the two groups (healthy and sick) were statistically different. Figure 5(a) shows the confusion matrix, and Fig. 5(b) shows the scatter plot results for healthy vs. diabetic feet. The confusion matrix shows the model’s sensitivity (93.3%) and specificity (100%). The mean, standard deviation (std) and standard error (SE) are listed above the plot in Fig. 5(b).
Next, we assessed the optimal camera parameters with respect to frame rate and exposure time. The merit function was defined as the influence on the accuracy of the model since this is the key marker in differentiating healthy from diseased limbs. The analysis was done numerically, by increasing the number of averaging frames, thus creating a larger de-facto camera integration period. The frames were averaged within the limit of the camera exposure time, which is bounded by its 1/frame rate.
An examination of the trends denoted by the red arrows in Table 2 clearly shows that the frame rate decrease had a negligible effect on model accuracy, suggesting that fast cameras may not be necessary. Figure 6 shows a typical comparison between a fast frame rate (840 fps) and its x10 fps reduction (84 fps). The Dynamic Laser Speckle was implemented in this example, and the plots were normalized to be at the same scale. The AUC is still clearly visible.
Table 2. Dynamic Laser Speckle: classification model accuracy vs. frame rate (fps) and exposure time analysis
In terms of exposure time analysis, we recommend using the shortest time possible. Note that increasing the camera’s integration time can increase the Signal-to-Noise ratio (SNR) in terms of photon flux, but also influences the smearing of the frames during ACF analysis, leading to less accurate model results. In some cases, the fps decrease may appear to improve model accuracy, e.g., 108 fps vs. 120 fps, for an exposure time of 7 [frames], but this is due to an approximation error, and does not change the overall trends described above.
A similar analysis as depicted in Table 2 was conducted for the LASCA method. A model accuracy of ∼87% was obtained for almost the entire fps span, starting from 840 fps, and decreasing to as low as 10 fps. Exposure was kept as short as possible (1/840 [sec], 1 frame). At low fps values (<40 fps) the accuracy of the Dynamic Laser Speckle based model started to decay, dropping to ∼65%, and below (Fig. 7(a)). This is reasonable, since the latter method is based on the decay of ${\textrm{g}^2}(\mathrm{\tau } )$, the ACF. When the sampling frequency drops to low values, there is a weaker association between decorrelation and changes in blood flow. By contrast, for the LASCA, a single frame was able to integrate all the speckles over the camera exposure time, independently of its previous frames. Figure 7(b) depicts model accuracy as a function of exposure time for a single frame rate of 30 fps. In LASCA we expected to have an optimal exposure time, usually ∼5-10 times longer than the decorrelation time [25]. A frame rate of 30 fps was chosen to mimic a typical low-cost camera. This analysis thus confirms the advantages of a simple method for exposure time optimization.
The relatively low speckle-to-pixel ratio results in fast decorrelation, and in turn decrease the SNR with respect to the suggested low frame rates. However the dynamics of the occlusion test alone are sufficient to achieve good differentiation results. Thus in this case high system SNR is the tradeoff for a more accessible solution, in terms of cost.
To track the physiological phenomena, the data must be sampled at the Nyquist frequency or higher, which corresponds to the fastest change. As seen in Fig. 3, the most abrupt change in 1/τ graph was at the cuff-release region, which was on the order of ∼2-3 [sec]. Thus, it was expected that the LASCA method would yield good classification accuracy, even with fps values as low as 10. A camera frame rate (fps) of 10 fps was found to be sufficient.
4. Conclusion
In this clinical study, we presented one possible application of a remote optical measurement configuration based on the analysis of back-reflected speckle patterns generated when illuminating tissue with a laser light. We analyzed the data with two well-known methods: Dynamic Laser Speckle and LASCA. We used an occlusion test to build a model to differentiate between healthy and diseased limbs. At frame rates typical of mid-range cameras (up to 1k fps), the analysis showed that both methods could be used as a basis for accurate classification modeling, at almost the same level of performance (∼ 90% accuracy). Furthermore, we demonstrated the advantage of LASCA in terms of its lower fps values and showed it could be decreased to as low as 10 frames per second, which is suitable for commercial cameras, and still maintain high classification accuracy (> 90%). An optimization of exposure time was done to increase LASCA performance.
In the measurements we were able to transfer the dynamics of the speckle patterns into a real-time estimation of a physiological quantity in the real world. We used low-cost commercial off-the-shelf components to demonstrate the robustness and ease of use of the sensor while still maintaining high accuracy in identifying diabetic patients vs. healthy controls. Thus the system cost could be reduced from hundreds of dollars for industrial fast cameras, to tens of dollars, for commercial common cameras.
Disclosures
The authors declare no conflicts of interest.
Data availability
The data supporting the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
References
1. R. O. Forsythe and R. J. Hinchliffe, “Assessment of foot perfusion in patients with a diabetic foot ulcer,” Diabetes Metab Res. 32 Suppl (1), 232–238 (2016). [CrossRef]
2. W. Karnafel, J. Juskowa, R. Maniewski, A. Liebert, M. Jasik, and A. Zbiec, “Microcirculation in the diabetic foot as measured by a multichannel laser Doppler instrument,” Med. Sci. Monit. 8(7), MT137–MT44 (2002).
3. J. E. Tooke, J. Ostergren, and B. Fagrell, “Synchronous assessment of human skin microcirculation by laser Doppler flowmetry and dynamic cappillaroscopy,” Int J. Microcirc Clin Exp. 2(4), 227–284 (1983).
4. M. F. Meyer and H. Schatz, “Influence of metabolic control and duration of disease on microvascular dysfunction in diabetes assessed by laser Doppler anemometry,” Exp Clin Endocrinol Diabetes 106(05), 395–403 (1998). [CrossRef]
5. K. F. Ma, S. F. Kleiss, R. C. L. Schuurmann, R. P. H. Bokkers, Ç. Ünlü, and J.-P. P. M. De Vries, “A systematic review of diagnostic techniques to determine tissue perfusion in patients with peripheral arterial disease,” Expert Rev. Med. Devices 16(8), 697–710 (2019). [CrossRef]
6. M. Kaspar, S. Partovi, M. Aschwanden, S. Imfeld, T. Baldi, H. Uthoff, and D. Staub, “Assessment of microcirculation by contrast-enhanced ultrasound: a new approach in vascular medicine,” Swiss Med Wkly. 145, w14047 (2015). [CrossRef]
7. A.W. Pollak and C.M. Kramer, “MRI in Lower Extremity Peripheral Arterial Disease: Recent Advancements,” Curr Cardiovasc Imaging Rep 6 6(1), 55–60 (2013). [CrossRef]
8. L. Potier, C. Abi Khalil, K. Mohammedi, and R. Roussel, “Use and Utility of Ankle Brachial Index in Patients with diabetes,” European Journal of Vascular and Endovascular Surgery 41(1), 110–116 (2011). [CrossRef]
9. F. Ezio, C. Giacomo, C. Maurizio, Q. Antonella, C. Vincenzo, and S. Francesco, “Evaluation of feasibility of ankle pressure and foot oxymetry values for the detection of critical limb ischemia in diabetic patients,” Vasc Endovascular Surg 44(3), 184–189 (2010). [CrossRef]
10. C. Cina, A. Katsamouris, and J. Megerman, “Utility of transcutaneous oxygen tension measurements in peripheral arterial occlusive disease,” J Vasc Surg 1(2), 362–371 (1984). [CrossRef]
11. F. Manfredini, N. Lamberti, V. Ficarra, E. Tsolaki, S. Straudi, P. Zamboni, N. Basaglia, and V. Gasbarro, “Biomarkers of Muscle Metabolism in Peripheral Artery Disease: A Dynamic NIRS-Assisted Study to Detect Adaptations Following Revascularization and Exercise Training,” Diagnostics 10(5), 312 (2020). [CrossRef]
12. S. P. Philimon, A. H. Kah Ching, and X. Ngu, “Optical measurement of tissue perfusion in diabetic wound patients,” in Bioengineering Principle and Technology Applications2 (Penerbit UTHM, 2019).
13. A. Nouvong, B. Hoogwerf, E. Mohler, B. Davis, A. Tajaddini, and E. Medenilla, “Evaluation of diabetic foot ulcer healing with hyperspectral imaging of oxyhemoglobin and deoxyhemoglobin,” Diabetes care 32(11), 2056–2061 (2009). [CrossRef]
14. M. A. Davis, L. Gagnon, D. A. Boas, and A. K. Dunn, “Sensitivity of laser speckle contrast imaging to flow perturbations in the cortex,” Biomed. Opt. Express 7(3), 759–775 (2016). [CrossRef]
15. D. A. Boas and A. K. Dunn, “Laser speckle contrast imaging in biomedical optics,” J. Biomed Opt. 15(1), 011109 (2010). [CrossRef]
16. Y. Li, R. Liu, Y. Wang, D. Wen, L. Meng, J. Lu, and P. Li, “Detecting relative speed changes of moving object through scattering medium by using wave front shaping and laser speckle contrast analysis,” Opt. Express. 24(8), 8382–8390 (2016). [CrossRef]
17. C. Ruangsetakit, K. Chinsakchai, and P. Mahawongkajit, “Transcutaneous oxygen tension: a useful predictor of ulcer healing in critical limb ischaemia,” J Wound Care 19(5), 202–206 (2010). [CrossRef]
18. C. J. Hauser, “Use of a transcutaneous PO2 regional perfusion index to quantify tissue perfusion in peripheral vascular disease,” Ann Surg 197(3), 337–343 (1983). [CrossRef]
19. M. Golberg, R. Califa, J. Garcia, and Z. Zalevsky, “Analyzing the requirements of high-speed camera parameters for enhanced laser speckle sensing of flow dynamics,” Eng. Res. Express. 2 (3), 035032 (2020). [CrossRef]
20. D. D. Postnov, J. Tang, S. E. Erdener, K. Kilic, and D. A. Boas, “Dynamic light scattering imaging,” Sci. Adv. 6(45), eabc4628 (2020). [CrossRef]
21. S. Katsui, Y. Inoue, K. Igari, T. Toyofuku, T. Kudo, and H. Uetake, “Novel assessment tool based on laser speckle contrast imaging to diagnose severe ischemia in the lower limb for patients with peripheral arterial disease,” Lasers Surg. Med. 49(7), 645–651 (2017). [CrossRef]
22. S. Kikuchi, K. Miyake, Y. Tada, D. Uchida, A. Koya, Y. Saito, T. Ohura, and N. Azuma, “Laser speckle flowgraphy can also be used to show dynamic changes in the blood flow of the skin of the foot after surgical revascularization,” Vascular 27(3), 242–251 (2019). [CrossRef]
23. J. W. Goodman, “Some effects of target-induced scintillation on optical radar performance,” Proc IEEE 53(11), 1688–1700 (1965). [CrossRef]
24. J. C. Ramirez-San-Juan, E. Mendez-Aguilar, N. Salazar-Hermenegildo, A. Fuentes-Garcia, R. Ramos-Garcia, and B. Choi, “Effect of speckle/pixel size ratio on temporal and spatial speckle-contrast analysis of dynamic scattering systems: Implications for measurements of blood-flow dynamics,” Biomed. Opt. Express. 4(10), 1883–1889 (2013). [CrossRef]
25. S. Yuan, A. Devor, D. A. Boas, and A. K. Dunn, “Determination of optimal exposure time for imaging of blood flow changes with laser speckle contrast imaging,” Appl. Optics 44(10), 1823–1830 (2005). [CrossRef]
