Automatic stent and catheter marker detection in X-ray fluoroscopy using adaptive thresholding and classification

Introduction

Intracranial stenting is one of the minimally-invasive therapies for neurovascular pathologies such as intracranial aneurysms and is accomplished via continuous X-ray imaging. Intracranial stents and flow diverters are small flexible mesh tube devices. Precise real-time device placement depends on the proper intra-operative visualization of the stent. Although radio-opaque stent markers are visible in fluoroscopy, they are as small as possible and their detection is challenging.

Sabrina et al. [1] proposed a technique to detect stent-graft markers to support the physicians with the stent-graft deployment in intra-operative fluoroscopic images. Fast radial symmetry transform following an ellipse or line fitting approach was used to detect a group of markers.

Movassaghi et al. [2] proposed a 3D representation of the deployed stent during a percutaneous transluminal coronary angioplasty procedure. Markers on the balloon catheter were utilized for motion compensation. Here the markers are clearly visible because of the high absorption coefficients and are detected using a scale-space algorithm.

Bismuth et al. [3] presented a technique to improve the stent visualization on an X-ray image sequence. Stent motion is inferred based on the motion of balloon marker balls as well as a guide wire. Potential marker balls are detected by local minima selection and a priori knowledge.

Wang et al. [4] proposed a fully balloon marker pair tracking and guide wire localization to employ 3D motion compensation and stent reconstruction. Marker candidates were detected using a probabilistic boosting tree classifier.

Many studies used blob-like structure detectors to detect markers [4], [5], [6]. Blob analysis actually refers to methods aiming at detecting regions with different characteristics, such as brightness or color, compared to their neighborhood.

Many approaches for stent marker detection are not applicable to intracranial stents because of different properties. Most studies focus on coronary stent enhancement (e.g., balloon-mounted stent with a couple of radio-opaque markers connected by a guide wire [3], [4], [5], [6]). This specific geometry is used to remove false positive candidate markers [6]. However, we are dealing with a type of intracranial stent, e.g., self-expanding stents with a different number of radio-opaque markers, which are not connected by a specific landmark.

In this paper, we focus on the X-ray opaque marker detection in intracranial fluoroscopic images. We have implemented a multi-step marker detection algorithm for fluoroscopy sequences based on an adaptive thresholding technique and binary classification.

We use blob analysis as feeding features to train classifiers. Different classifiers include linear discriminant analysis classifier (LDA), naïve Bayesian classifier (NB), k-nearest neighbors (KNN) and support vector machine (SVM) which will be described in the next section.

Methodology

We develop an algorithm to detect radio-opaque markers located on intracranial stent and catheter.

Experiments and results

The framework for our approach is illustrated in Figure 1. For training and test, we used four and three clinical fluoroscopy sequences, respectively. To detect true X-ray markers, classifiers are trained to discriminate marker regions from non-marker regions. Features were extracted from 112 patches of some arbitrary frames of 4 clinical fluoroscopy sequences to train the classifiers. Different classification techniques were employed to differentiate patches. SVM classification was performed by standardizing the features, and sequential minimal optimization (SMO) was chosen as an optimization routine. The radial basis function was selected as a kernel function. For the KNN classifier, the distance metric was Euclidean distance and the number of nearest neighbor (K) was 1. Both LDA and NB classifiers were trained in their default mode in MATLAB (MathWorks, Natick, USA). To evaluate the performance of the approach, 1948 test patches were extracted from various frames of 3 different fluoroscopy sequences. The number of true positive (TP), true negative (TN), false positive (FP) and false negative (FN) patches were calculated for each test frame. TP and TN values indicate the number of patches that were truly classified as patches with markers and patches excluding markers (including other objects other than radio-opaque markers), respectively. Different criteria including sensitivity, specificity and precision were used to assess the performance of different classifiers.

Figure 1:

Workflow for the proposed approach.

The results are illustrated in Figure 2 for the first test data set with the four classifiers: KNN, NB, SVM and LDA. Figure 2a shows the original input frame, stent and catheter are specified on the image. Figure 2b is a pre-processed frame that is used for thresholding. X-ray opaque markers can be better visualized on the pre-processed image. Figure 2c, shows the candidate regions for markers that are surrounded by a green circle and are shown as white regions within the image. Finally, Figure 2(d–f) shows prediction results for KNN, NB, SVM and LDA classifier, respectively. Regions marked with green and red stars illustrate predicted marker locations and predicted non-marker areas, respectively. In this specific frame, KNN predicts two markers truly, while it misses 5 other markers. NB has the best performance for this specific frame by 6 TP, 1 FN and 1 FP. SVM predicts 3 markers truly while it misclassified 4 other patches (4 FN). LDA also misclassified three marker patches.

Figure 2:

The results of classifications for the first test dataset, a. Original input image, b. pre-processed image, c. Detected areas after thresholding as possible marker candidates, Classification result by d. KNN, e. NB, f. SVM, g. LDA.

The performance of classifiers on 3 test sequences was compared using different criteria (See Table 1). Based on the accuracy results, the NB classifier performs the best on the first test sequence with 72.29%. In the second test sequence, SVM has the highest accuracy of 85.83%. In the last sequence, LDA and SVM have almost the same accuracy value with 97.61 and 96.97%, respectively. Based on the sensitivity value for different classifiers, it concluded that the NB classifier has a lower number of false negatives. NB has the highest mean sensitivity value with 87.76% and in general based on accuracy, specificity and precision SVM works the best and its average values are 81.56, 91.94 and 84.21%, respectively.

Conclusion

Markers are very small and hardly visible in 2D X-ray fluoroscopy and our proposed approach can be used to assist clinicians during the intervention with the stent location in 2D by highlighting the detected markers on X-ray images. This eventually leads to faster procedures with less dose exposure to both patients and clinicians. Furthermore, detected markers may be helpful in improving stent visibility in 2D. The adaptive thresholding method used in this study can detect even very small and hardly visible markers, however, to detect all possible candidates, it still depends on the specified sensitivity. Increasing the sensitivity will increase the number of false positives that have to be ruled out in the classification step. This emphasizes the importance of the classification step. Different classifiers including KNN, NB, SVM and LDA were used as a binary classifier to discriminate between non-marker areas and marker regions. Based on the evaluations, SVM outperforms other classifiers in terms of accuracy, specificity and precision. The highest specificity for SVM with 91.94% was obtained at the cost of low sensitivity. Furthermore, based on the obtained accuracy values for different classifiers, it can be concluded that the performance of the classifiers depends on the dataset. Possible improvements include increasing the number of training patches and increasing features and employing cross-validation to improve the classifiers’ generalizability. Since stent and catheter markers are different in terms of size and shape, employing multi-class classification and considering different types of markers may improve the performance of the classifiers.