The segmentation methodology is based on combining operations in three steps, Preprocessing, Feature Extraction and Binary Morphological Image Reconstruction (Figure 1). The evaluation was performed comparing the segmented images with their gold standards made by experts and calculating the parameters of accuracy [15, 32]. The material is composed by a set of 290 IVOCT images from 2 patients, 2 pigs, and 1 rabbit, from the database of the Heart Institute of the University of São Paulo Clinic Hospital, Brazil (InCor). The 290 images in the dataset were chosen to represent a variety of coronary feature in IVOCT images, such as different degree of wall contrast, lumen irregularities due to thrombus, plaques and branches, and with 30 and 180 days after stent implantation, the study protocol was approved by the ethic committee of InCor with informed consent signed by patients. The images were acquired with pullback of 0.5 mm/s, and 20 f/s, by a TD-OCT, St. Jude/LightLab ImageWire catheter, connected to the St. Jude/LightLab OCT Imaging System and Probe Interface Unit (St. Jude/LightLab Optical Coherence Tomography – St. Jude Medical, Inc., Westford, Massachusetts, USA).
Preprocessing
Because, cardiac centers have images acquired and stored in the usual format for visual evaluation, Cartesian domain with catheter reflection and alignment marks, the preprocessing block should prepares and normalizes the image, providing a standard image to the rest of the method. If OCT raw data was available, the preprocessing block could be neglected; however, since we could not ensure that all previous acquired was exported and saved in this format, the preprocessing is necessary. Therefore, beyond image normalization, this stage also aims at the attenuation and enhancement of undesirable and desirable features, respectively [4, 42]. Specifically, the catheter reflection, and the alignment mark are undesirable features for this purpose, and may damage and limit the segmentation procedure. On the contrary, work with circular structures, such as coronary, in the polar domain has many advantages due to its 1D appearance [4].
The catheter reflection, and the alignment marks are recognized by a ring at the center of the IVOCT image, and straight lines marking fixed positions, and one long line crossing the image (Figure 2a). For our purpose, they can be seen as noise, hence dropping down the segmentation accuracy, because they may be misinterpreted as tissue during the Feature Extraction procedure. However, since they have known location, dimensions and characteristics, they can be removed by two simple operations. First, the catheter is removed by eliminating the concerning pixels inside the catheter ring maximum radius (r
Max
) (Figure 2b) [32]. Secondly, a 2D median filtering procedure, using 5 by 5 window, was carried out to attenuate the alignment marks, and also fading out any destructive Speckle effects without damaging borders [43, 44] (I
Filtered
) (Figure 2c).
Working in an appropriate domain may help improve the method efficiency, and simplify image description [4, 32]. Because the coronary has circular structure in the Cartesian image, a 1D appearance is obtained when converted to the polar domain. Therefore, so as to facilitate next procedures [7], the images were transformed into the polar representation (I
PreProc
(r, θ)) (Figure 2d), with 200 pixels of r, equal the length of the Cartesian image radios, and 630 pixels of θ, approximately equivalent to a radial variation of 0.57 degrees per line. These dimensions are important because further morphological procedures uses operations based on I
PreProc
(200, 630).
Feature extraction
The Feature Extraction uses operations to identify and to distinguish the desired information; hence, increasing discrimination and improving classification [42, 45, 46]. Following what was successfully applied in [32], a combination of two widely used operations, Discrete Wavelet Packet Frame (DWPF) [42, 47] and Otsu threshold [48], were adopted to acquire tissue information (Figure 3).
The Discrete Wavelet Packet Frame (DWPF) is well known and has been established as a very important tool to distinguish the desired information from others, hence increasing the separability between them [4, 32, 42, 47, 49, 50]. Therefore, one level of decomposition using Daubechies 1 (dB1) was carried out [32, 49, 50], and the I
PreProc
image (Figure 4a) was decomposed into four coefficients (Figure 4b). The wavelet and decomposition coefficient were selected, based on the high correlation with the tissue information. As can be seen in Figure 4c, the Coefficient of Approximation 1, cA1, is the one that best extracted and separate tissue information (Figure 4c, between yellow to red color) from the rest of the image. Once we have the tissue information, the lumen region is directly recognized (Figure 4c). Therefore, cA1 was chosen to be the tissue information supplier, hence serving as reference for the binary lumen object reconstruction. Binary morphological image reconstruction [32, 51] is a very useful tool to estimate and polish previous information, thus increasing the method accuracy and robustness. However, a binarization process should be performed beforehand. Due to the variety of resultant IVOCT image features, according to the artery and the patient being imaged, an adaptive threshold selection is required for a good binarization.
Otsu [48] is a dynamic threshold selection method for dynamic binarization process, in which a histogram is divided into two classes, by seeking for the smallest variance between two clusters, hence providing a good separation for data with bimodal histogram. Because the wavelet transformation increases the separability of desired and non-desired information, a highly bimodal histogram is created with cA1, which makes an adequate data to be binarized by Otsu. However, since infrared is distance sensitive, the contrast between tissue and blood may have an angular intensity variation according to the catheter location (Figure 5a, red square). Consequently, data between the two classes may appear in the histogram (Figure 5b), highlighted in black); hence, information may be lost after binarization (Figure 5c, highlighted in red). Nonetheless, the histogram of each column of the cA1 usually has two pieces of information (Figure 5d and e), tissue and no-tissue; even when the tissue contrast is low, a bimodal histogram will be obtained (Figure 5e, column b). Therefore, we adopted a local Otsu binarization process, by column; consequently, by performing the mentioned procedure, cA 1bin is created, which corresponds to the binary version of cA1 (Figure 5f). As a result, the Binary Morphological Reconstruction can be carried out.
Binary morphological image reconstruction
Binary morphological Image reconstruction is a sequence of combined mathematical morphology techniques [32, 51, 52] designed to obtain an accurate binary version of the desired object. Particularly in this block, we used the previous information, cA1
bin
(Figure 6a), to obtain the corresponding binary lumen object, l
bin
(Figure 6b). In order to accomplish this task, the operations for the reconstruction are divided into three parts, Polar Image Reconstruction, Opening Detection and Correction, and Cartesian Image Reconstruction (Figure 6). The first, polar image reconstruction, aims to obtain the complete complementary part of the polar lumen object, l
polar
*, (Figure 7). If the image has branch opening, the opening detection and correction block is performed to correct it (Figure 8). The final binary lumen object, in the Cartesian domain, is reconstructed during the Cartesian image reconstruction (Figure 9). Each block is detailed below:
Polar image reconstruction
Polar Image Reconstruction is a combination of binary morphological procedure applied in the polar domain information, cA1
bin
, so that it can be refined, and possible missing information estimated (Figure 7). Due to the range of artery and blood features of patients, spurious noises may appear in a variety of sizes and quantity in the lumen region in the cA1
bin
(Figure 7a). Since they could also be connected to the tissue information, these noises must be removed. To remove them, we first disconnect them from the main tissue information block by filtering the image with a morphological opening procedure [51, 52] resulting in cA1
binFiltered
(Figure 7b); second, an upward filling procedure [32, 51–53], resulting in cA1
binFilled
(Figure 7c), followed by an area selection generating the selected(cA1
binFilled
) is carried out; finally, a last closing procedure [51, 52] is performed, obtaining the complementary polar lumen object, l
polar
* (Figure 7d). The opening and closing procedures uses circular structuring elements, with 3-pixel and D pixels diameters, respectively. The circular elements is to maintain the smooth contour of object, and D = r
Max
pixels, is an adaptive diameter where r
Max
correspond to the catheter ring maximum radius [32]. This size assures that possible lumen border irregularities will be attenuated without changing original contour or connecting the object to top of the polar image.
Opening detection and correction
Branch openings are shadows in IVOCT images caused by vessel bifurcations during image acquisition (Figure 8a) [32]. Consequently, they are propagated to the preprocessed image and cA1 (Figures 8b, and c). Because the gap does not produce contrast in its columns, a bi-modal histogram is not generated; thus, the columns corresponding to the gap are binarized as level “1” (l
polar
*
Opened
) (Figure 8d). This causes high derivative at the lumen border shape of the l
polar
*
Opened
(Figure 8d). Therefore, its detection and correction is carried out as follows. First, a signal representation of the polar image is created h(l
polar
*
Opened
) (Figure 8e). Second, the signal derivative is calculated, and by finding values higher than a threshold, the opening is detected. Consequently, the correction initiates by removing all the values corresponding to the gap (Figure 8f), and performing Piecewise cubic Hermite interpolation (Figure 8g). Finally, the corrected polar image (l
polar
*
Corrected
) (Figure 8h) is then reconstructed using the interpolated signal. The beginning and end of the gap are identified as the first and last derivative absolute values, respectively higher than the threshold, which is defined as 5 standard deviation of all derivative signals.
Cartesian image reconstruction
The Cartesian image reconstruction combines an image domain transformation with one last morphological operation, for object polishing. Therefore, l
polar
* (Figure 9a) is obtained, no matter if it went through the opening correction. First the logic negation of l
polar
* is carried out; hence, obtaining the lumen object in the polar domain, I
polar
(Figure 9b). Finally, the lumen reconstruction, lumen
bin
(Figure 9c), is concluded by transforming to the Cartesian domain, I
Cartesian
followed by one last opening operation [51, 52], with a circular structuring element with adaptive diameter Rmin pixels (S
circ(Rmin)
), where Rmin is the minimum radius between the center and border of the lumen. This last opening is because possible irregularities in the polar domain will be carried to Cartesian. Using a circular element, these remaining irregularities are removed, and a smooth contour of the object is obtained. Finally, the segmentation is concluded by extracting and placing contour of lumen
bin
on the Original image [52] (Figure 10).