Intelligent epidural needle placement using fiber-probe optical coherence tomography in a piglet model.

Incorrect needle placement during an epidural block causes medical complications such as dural puncture or spinal cord injury. We propose a system combining an optical coherence tomography imaging probe with an automatic identification algorithm to objectively identify the epidural needle-tip position and thus reduce complications during epidural needle insertion. Eight quantitative features were extracted from each two-dimensional optical coherence tomography image during insertion of the needle tip from the skin surface to the epidural space. 847 in vivo optical coherence tomography images were obtained from three anesthetized piglets. The area under the receiver operating characteristic curve was used to quantify the discriminative ability of each feature. We found a combination of six image features-mean value of intensity, mean value with depth, entropy, mean absolute deviation, root mean square, and standard deviation-showed the highest differentiating performance with the shortest processing time. Finally, differentiation of the needle tip inside or outside the epidural space was automatically evaluated using five classifiers: k-nearest neighbor, linear discriminant analysis, quadratic discriminant analysis, linear support vector machines, and quadratic support vector machine. We adopted an 8-fold cross-validation strategy with five classifications. Quadratic support vector machine classification showed the highest sensitivity (97.5%), specificity (95%), and accuracy (96.2%) among the five classifiers. This study provides an intelligent method for objective identification of the epidural space that can increase the success rate of epidural needle insertion.


Introduction
Epidural blockade is an effective anaesthesia and analgesia technique which is widely applied in neuraxial anaesthesia, labour analgesia, and acute and chronic pain therapy. Epidural blockade procedures are estimated to account for 10% of all anaesthetic procedures worldwide. The epidural space (ES) is located between the ligamentum flavum (LF) and the dura mater and has a thickness ranging from 2 to 6 mm. The loss of resistance (LOR) method with air or saline is the standard method for epidural administration [1]. However, LOR is highly dependent on the operator's experience, and learning to place the needle competently requires practice over an average of 60-90 epidural placement procedures [2]. The epidural technique using LOR is associated with a total failure rate of up to 10% [3,4]. Needle positioning is even more challenging in patients with overweight or structural spinal abnormalities. Investigators have used optical reflectance [5,6], optical spectrum [7], ultrasound [8], and nerve stimulation [9] to improve the accuracy of needle placement.
In a previous study [10], we established a prototype that uses a guided epidural block method by combining a rotated side-looking fiber probe with swept-source optical coherence tomography (SSOCT). This prototype provides a series of visual images and thus facilitates needle positioning by identifying the structures of the different tissues that surround the needle tip. Our method presented the first such demonstration of "live" two-dimensional (2D) images, which can assist navigation toward the ES [10]. Although this technique shows high sensitivity and specificity, the high inter-expert variation makes decisions greatly divergent. Therefore, an objective judgment tool that can help physicians evaluate the OCT images is required.
Feature analysis has become an important method for medical image classification [11][12][13]. Most of the classifiers commonly applied to image data were originally trained in statistics and machine learning. All classifiers use a band of breeding data to define the decision boundaries in the data space, although different classifiers use different decision boundaries. For example, hyperplanes are usually used as boundaries in linear classifiers, while more complex nonplanar boundaries are used in nonlinear classifiers. Additionally, different classifiers have different methods of boundary placement from training (e.g., a linear support vector machine [SVM] places the decision hyperplane such that the margin to the patterns is maximized on either side).
Therefore, we propose an automatic identification (AI) system, an objective method to solve the problem of variations in needle-insertion technique. In this approach, OCT images were continuously acquired during insertion of the needle tip from the skin surface to the epidural space. Eight image features were extracted from OCT images collected from various tissue structures. To objectively determine whether the needle position is inside or outside the ES, inputting eight features extracted from the OCT training images group constructed intelligent classifiers. Five classifiers, including quadratic support vector machine (QSVM) [14], linear support vector machines (LSVM) [15], quadratic discriminant analysis (QDA) [16], linear discriminant analysis (LDA) [17], and K-nearest neighbor (KNN) [18], were tested and compared to assess their ability to classify OCT images and thereby allow automatic needle-tip positioning.

Animal preparation and epidural insertion
Our animal study was approved by the Institutional Animal Care and Committee of Taipei Veterans General Hospital (Taipei, Taiwan). Chinese native pigs with an average weight of 25 kg were studied. The animals were intubated following induction of general anaesthesia with tiletamine/zolazepam 50 (Virbac, France). Anaesthesia was maintained with isoflurane under mechanical ventilation. The piglets were placed in the left lateral position for epidural placement. The SSOCT imaging system is shown in Fig. 1(A). The light source is a high-speed swept laser (Axsun Technologies Inc, Billerica, MA) with a central wavelength of 1310 nm, a tuning bandwidth of 110 nm, and a sweep rate of 100 kHz. The system resolutions in the air were ~11 µm axially and ~15 µm laterally. The main output of the laser is coupled to a fiberbased Michelson interferometer through a circulator and split into the reference and sample arms using a broadband 50/50 coupler. The sample arm of the SSOCT system was constructed with a single-mode fiber and the needle probe, and was connected to a distal rotary motor. Finally, the reflected light propagating backward through the coupler and circulator was simultaneously guided onto a balanced detector and recorded using a highspeed analogue-to-digital (A/D) converter (Alazar Technologies Inc., Quebec, Canada). Realtime differentiation of the needle tip inside or outside the ES was automatically evaluated by the AI system.
The design of our fiber needle probe is the same as our previous publication [10], demonstrated in Fig. 1(B), was constructed by attaching a single-mode fiber to a gradient index lens and a prism (BK7 Aluminum coated) and was protected by covering an optical glue in a plastic catheter with a 0.9-mm outer diameter. Thus, the output light has a fixed deviation angle of nearly perpendicular to the probe's axial direction. In this experiment, an 18-gauge insulated Tuohy needle (Teleflex Incorporated, Morrisville, NC, USA) was used as a tunnel to allow the OCT probe to reach the ES. A "locking holder" in Fig. 1(B) was designed to keep the probe tip within the Tuohy needle during the rotation to protect the optical probe. Thus, the offset between image location and the needle tip was measured to be around 1mm. When a 2D OCT image was acquired by the circumferential scanning of the optical probe with a rotational motor, this probe design provides a "side-looking" image "near" the needle tip, and the imaging location is in perpendicular to the probe axis.
In our previous reports [6,10], a linear mixed model analysis with compound symmetric covariance structures for repeated measures has been used to evaluate the interclass and intraclass correlation. The optical properties of inside and outside of ES tissues were compared at different puncture sites in each measurement from the same pig, and in measurement from different pigs. The statistical results indicated that there was no significant difference in the optical properties obtained among the six insertions of five piglets. Therefore, in this study, the needle probe was inserted from the skin hypoderm, through muscle, LF, and then into the ES in three piglets at different spinal levels, from the thoracic to the lumbar regions, using a paramedian approach. By circumferentially scanning the OCT probe with a rotational motor, continuous 2D OCT images (r-θ polar coordinates) were acquired as the needle was progressively inserted into the skin surface toward the ES. The image was displayed at a rate of 80 frames per second. Each frame consisted of 1000 A-lines, with an imaging depth of approximately 5 mm in the air. Figure 1(C) shows representative tomograms of different structures as the needle moves from the muscle toward the ES in one insertion path. Because there is a distance of 1 mm between the fiber tip and the needle tip (as demonstrated in Fig. 1(B)), the needle tip may reach to the ES when the OCT image shows LF structures (i.e. in the case of our piglet study, LF thickness is smaller than 1mm). Thus, in this study, OCT images showing only muscle tissue indicated that the needle tip was positioned outside the ES, and the needle advancement was stopped when observing either the epidural fat tissue or LF structures in the OCT images, based on the criteria listed in our previously published report [10]. The conventional loss of resistance (LOR) technique was then used at the same time to confirm the position of the needle tip is within ES. Once the ES region was suspected, we opened the "locking holder" and pushed the fiber probe beyond the needle tip by approximately 10 cm and obtain a circular image within the ES to double-check the ES region. As similar to what established in our previous work [19], adjacent frames from each continuous rotation could be stacked to show a 3D volume of the epidural space. The occurrence of these events means that the needle is correctly placed inside the ES, as successively confirmed by using radiograph (KXO-50R; Toshiba, Tokyo, Japan) with 5-ml contrast (ioxitalamic acid). The animals were euthanized after this procedure.  2) 4. Otsu's method was adopted, which provides an automatic threshold to minimize the intra-class variance of black and white pixels while maximizing the distance between the mean values of black and white pixels. This global threshold (i.e., level) can be applied to convert an intensity image to a binary image. Thus, we can compute the grey level ratio as the percentage of white pixels occupied in each 2D OCT image.
These eight features describe the distribution of pixel values of the image intensity within each OCT image through commonly used calculations and basic metrics. N is the total number of pixels in an image, and I is the intensity in each pixel. Because there is a distance of 1 mm between the fiber tip and the needle tip, and the LF thickness in piglets is about 1 mm in our experiment, the needle tip is within the ES when the OCT image shows LF structures. Thus, OCT images showing either LF or ES structures indicated that the needle tip was positioned inside the ES, while OCT images showing only muscle tissue indicated that the needle tip was positioned outside the ES.

Performance evaluation and classification analysis
Receiver operating characteristic (ROC) curves were used to show the discrimination capability when each feature was used to discriminate between the inside of the ES (including the LF and ES) and the outside (muscle). The area under the ROC curve (AUC) was used to quantify the discriminative ability of the test.
To automatically and objectively determine that the needle position is inside or outside the ES, five classifiers were applied by inputting eight features extracted from OCT training group images. Figure 2 shows a processing flowchart where 7/8 of the training images were initially used for model construction, and the remaining image was used in a test set for validating the predictive performance. Once the model was constructed successfully, the classification line was generated before inputting the test group images. Iterations were then performed using a different test set until all the training images had been used. Thus, for validation of the predictive performance of each classifier, the sensitivity, specificity, and accuracy of test images by using different classifiers were evaluated and compared.   Figure 3 shows representative OCT images in three piglets. Signal-rich layered structure in the muscle tissue is due to backscattering from the muscle fascicles. The inclusion of dispersed adipose tissues also leads to non-uniform signal distribution. The light scattering from the LF is stronger than that in muscle; this may be because the LF tissue is mainly comprised of elastin, which is far denser than muscle tissue. Thus, the LF has strong and homogeneous signal distributions with a smaller imaging penetration depth than that of the muscle tissue. ES images also display a smaller imaging penetration depth of adipose tissue and show numerous empty spaces with no reflective light.    To objectively determine whether the needle position is inside or outside the ES, intelligent classifiers were constructed by inputting eight features extracted from the OCT training images group. Iterations were performed using 105 different images for each model test until all the 847 images had been used. The accuracy values of using all image features for five classifiers, QSVM, LSVM, QDA, LDA, and SVM, are shown in Table 1. The accuracy of the classifiers for determining the inside vs. outside of the ES by using all features of the images was as follows: QSVM, 96.7% (AUC, 0.967); LSVM, 96.7% (AUC, 0.967); QDA, 94.7% (AUC, 0.947); LDA, 97% (AUC, 0.97); and KNN, 92.5% (AUC, 0.925). The first, second, and fourth classifiers, i.e., QSVM, LSVM, and LDA, showed the highest sensitivity, specificity, and accuracy (>96%). (Represented with red font). Validation results using the six selected features (i.e., Mean, Mean_depth, Entropy, Mad, RMS, STD), which showed AUCs of 70%, classified by the five classifiers are shown in Table 2. The first classifier, QSVM, showed the highest sensitivity (97.5%), specificity (95%), and accuracy (96.2%). Figure 6(A) shows the process time for calculating each feature from one OCT image. Figure 6(B) compares the total processing time for extracting eight features (12.7 ± 4.2 ms) and six selected features (9.5 ± 3.0 ms). The standard deviation was calculated from ten iterations of each experiment.

Discussion
This study proposes an intelligent epidural needle placement technique using a fiber-probe SSOCT integrated with an AI system. The selection of features used as distinguishers in different tissues is crucial in the development of our AI system. As demonstrated in Fig. 3, the characteristic features visible in OCT images are similar in the same tissue of all three piglets. OCT images of muscle layer have a layered structure formed of fascicles and a larger tissue penetration depth than the LF or the ES. LF tissue images have strong and homogeneous signal distributions with a smaller imaging penetration depth. ES images also display a smaller imaging penetration depth of adipose tissue and show a broad range of space with no reflective light. According to these significant features of LF, ES, and muscle tissues, eight quantitative image parameters for tissue classification were extracted from each in vivo OCT image (Fig. 4). We found that the values of Mean, Mean_depth, Entropy, Mad, and RMS in LF and ES are smaller than that in muscle tissues. The ROC comparison curves also show that six features (i.e., Mean, Mean_depth, Entropy, Mad, RMS, STD) have distinguishability (AUCs > 0.7).
We then compared the ability of five classifiers (i.e., QSVM, LSVM, QDA, LDA, KNN) using eight or six image features to discriminate the inside of the ES from the outer tissues. After the training stage, each classifier provides the best boundary plane that divides the individual point data of the separate categories. Thus, once we completed the model construction, each test data point was mapped into the space determined by the training process so that the inside or outside category was predicted based on the side of the gap that point falls on. These automatic classifiers provide a real-time and objective method for ES identification. Among them, the QSVM classifier showed high sensitivity (97.5%, 97.5%), specificity (96%, 95%), and accuracy (96.7%, 96.2%) in tests using eight and six image features, respectively (Table 1 and Table 2). The LDA classifier also showed good performance (i.e., 99% sensitivity, 95% specificity, and 97% accuracy, Table 1) when using eight features, but showed 98% sensitivity, 91% specificity, and 94.5% accuracy ( Table 2) when using only six selected features.
In addition to the performance metrics described above, the ROC curve is another important measurement used to select the best classifier [20][21][22]. We compared the performance of five classifiers by calculating the AUC values, noting that the AUC value is always between 0 and 1 and thus the best classifier is the one with a relatively high AUC value. The results in Table 1 show that the AUC values for both QSVM and LDA (~0.97) are higher than those of the other classifiers. When only six features were used with the five classification methods, the AUC for QSVM (0.96) was the highest (Table 2).
In addition to high accuracy and objective classification, real-time ES identification is also essential. Although epidural needle placement depends on the doctor's experience, the needle usually spends at least 2 minutes traveling from the skin surface to the ES. In this study, a series of 2D OCT circumferential images was built up by continuously rotating the optical probe with a rotary motor while the needle was progressively inserted toward the ES. The OCT image was acquired and displayed at a rate of 80 frames per second, which is rapid enough for providing real-time imaging at the exact position around the needle tip. Features extraction for one OCT image required only 9.5 ms when using six selected features. This processing time was three-fourths of that noted with eight-feature extraction, with nearly equal sensitivity, specificity, and accuracy when using the QSVM classifier. Real-time classification by an AI system using different classifiers takes 11.8 ms (QSVM), 11.5 ms (LSVM), 7.1 ms (QDA), 9.7 ms (LDA), and 13.6 ms (KNN). Thus, the classification of one OCT image takes only 20 ms when using six selected features and the QSVM classifier, providing a real-time identification speed of 50 frames per second.
Nevertheless, our method has some limitations. First, the translation of our results to humans is limited by the noticeable anatomical differences between pigs and humans. For example, when our OCT image shows the LF structure, the needle position will be classified as the inside of ES by the AI program in this study. However, the tissue thickness of human LF is about 2~6 mm, which is much higher than the 1-mm thickness of the piglet LF [23]. For future human studies, the AI program will need to be modified such that LF images can be classified as tissue outside the ES. Further studies to validate these results in human participants are needed to provide data that will help this method achieve clinical utility.
Second, the image quality may be severely influenced by the presence of blood or blood clots around the needle tip because of its highly scattering property at a wavelength of 1310 nm. (e.g., Fig. 7(A) shows a dark central region with a high level of circular scattering in the OCT image when our fiber probe through a small vessel or microvascular in muscle tissues.) However, since the scattering and absorption in water are very low [24], injection of some water or saline solution can clear the imaging field of view. A washing device will be needed on our future probe design to efficiently remove highly scattering interference from the blood. Third, several misclassification situations occurred. Figure 7(B) includes multiple substructures, both LF and fat tissue were within one image. Figure 7(C) having a broad distribution of adipose fat tissue in the ES. These two cases were all misclassified into muscle tissues, which may due to their highly scattering features. Figure 7(D) to 7(F) show a relative weak scattering and shallow depth penetration in muscle tissues, which were all misclassified into tissues inside of ES. Therefore, features extraction by using small sub-image units (e.g., 10 pixels x10 pixels) will be considered in our future work, which can provide a regional classification for each different area and is expected to improve the classification accuracy. In summary, the present study demonstrates the ability of the combination of our proposed AI system and OCT imaging for automatically and objectively identifying the location of the needle tip. Only six quantitative image features, i.e., mean, mean with depth, entropy, Mad, RMS, and STD, were necessary for an AI system. Compared with the subjective physician decisions, this method provides higher accuracy. We anticipate that QSVM-based classification of OCT images will improve the accuracy of epidural placement and therefore reduce medical complications associated with the neuraxial blockade. Further clinical studies are needed to validate this preliminary animal study.

Disclosures
The authors declare that there are no conflicts of interest related to this article.