Hessian analysis for the delineation of amorphous anomalies in optical coherence tomography images of the aortic wall

Eusebio Real; José Fernando Val-Bernal; José M. Revuelta; Alejandro Pontón; Marta Calvo Díez; Marta Mayorga; José M. López-Higuera; Olga M. Conde

doi:10.1364/BOE.7.001415

1. Introduction

Aortic aneurysm is a cardiovascular disease affecting the wall of the aorta artery. Degradation of the wall consists mainly of the loss of smooth muscle cells and elastin fibers, inducing weakening of the vessel wall [1]. The main manifestation of the pathology is a vessel diameter increase, continuously undermined as the periodic heart pumping softens its structure. When the artery is too weak to stand this heart pump rushing, it fissures producing a massive blood leakage. Before getting to this point, surgical intervention is needed to prevent patient’s death. Surgical interventions consist on placement of an endovascular stent-graft or, when the pathology is severe, an external graft prosthesis by open surgery intervention. In the last case, the part of affected tissue is excised and replaced by a synthetic tube. The suture must be sewn to a healthy region of the remaining vessel or, otherwise, the aneurysm will reproduce or cause an aortic dissection.

The dimension of the areas of vessel wall affected by degradation is the real magnitude of degradation [1]. Hence, an accurate quantification of this magnitude is a hallmark of vessel health. To assess the actual condition of the artery wall, conventional staining techniques (such as Hematoxylin and Eosin or Verhoeff’s Van Gieson) combined with common microscopy allow the visualization of the status at a cellular level. However, these techniques are time consuming and difficult to be integrated into the surgical workflow. Moreover, since the aortic tissue is thick and difficult to handle, the histological evaluation is limited to specific sampling points, once the tissue is excised, losing the evaluation of overall spatial status of the vessel wall.

Optical Coherence Tomography (OCT) is an imaging technology that allows the visualization of tissue samples in depth, imaging few millimeters within the sample. Resulting OCT images from tissue are equivalent to those obtained with common histology techniques, with the advantages of continuous and fast acquisition, easy incorporation into the surgical procedure but at fairly less resolution. Previous work [2] shows how the OCT approach correlates with the information provided by common histological techniques. Therefore, OCT could provide real-time guidance information about the vessel state under intraoperative conditions if an accurate and automated delineation of degradation is developed.

For delineation purposes, Hessian analysis has been previously proposed to detect vascular vessels, with tubular shape, in two [3,4] and three dimensional [5] images obtained by Magnetic Resonance Angiogram (MRA). Further modifications of this method were developed to detect nodules and vessels, based on the detection of predefined lines and blob-like [6,7] or curvilinear structures [8]. Multiscale analysis [3] has been also used to analyze OCT images for the study and quantification of vessel shape in the case of lymphatic vessels [9], angiography in the retina [10] and in extremities [11] and for segmentation purposes in the case of pig airway wall [12].

In contrast with previous experiences of Hessian analysis, in the present study of degradation of the aortic wall, the main issues are the lack of predefined shape and size of the pathological regions and the presence of high levels of noise due to reflections in small inhomogeneities along the wall thickness. In addition, attenuation of the aortic tissue strongly reduces sensitivity within depth, reducing the dynamic range of deeper areas (in contrast with the more superficial ones) and blurring anomalies boundaries, making anomaly segmentation a challenging process [3, 13, 14]. An alternative Hessian analysis is proposed to delineate degradation areas of amorphous shape, different sizes and intended to work under noise and variable dynamic range conditions, as it occurs in OCT images of the aorta wall. A new delineation accuracy estimator metric is also proposed to evaluate and optimize the anomaly detection performance avoiding a subjective evaluation. Finally, the detected anomalies have been used to quantify degradation in the aorta and this degradation has been compared against its histological quantification.

2. Matherials and methods

2.1 Aorta samples and protocol

In this viability study, human ex-vivo samples were obtained from the thoracic ascending aorta (TAA) region. Specimens come from two groups for comparison: healthy donors (11 patients), in compliance with the standard heart donor criteria, and diseased aneurysm samples (21 patients) from aneurysm surgical open repair interventions. Specimens were conserved in Phosphate Buffered Saline (PBS) and measured with the OCT system within 12 hours from excision. Afterwards, they were conserved in a diluted formaldehyde solution upon histological analysis.

OCT measurements were taken at several regions on each specimen and 5 OCT images, i.e. B-scans, were taken on each regions (356 B-scans for the whole specimen set). The number of regions depends on the size of the specimen, being a number between 1 and 4, with dimensions of 10x10mm. During histology, several cuts were applied to the regions, considering only the quantification of the region presenting the highest degradation.

2.2. Histological analysis

Five different stains and procedures were applied to each sample: Hematoxylin and Eosin, Verhoeff’s Van Gieson, Alcian Blue at pH 2.5 and Alpha-smooth muscle actin [2]. The dimensions and severity of degradation in the aorta wall were measured by expert pathologists, providing the degradation state of the affected regions, quantified with a degradation histology score [1, 15], grading from null degradation (score 0) to severe degradation (score 15).

2.3 Optical coherence tomography setup

All samples were analyzed with the Thorlabs OCS1300SS swept source system. The working wavelength is 1325nm ± 50nm bandwidth. The sensitivity of the system is up to 100dB. Cross-sectional B-scan images have dimensions of 10mm width by 3mm depth in air, sampled with 1024 pixels width by 512 pixels depth. Intensity level was digitalized with 8 bits.

The imaging probe was focused on the intima surface of samples to obtain information of the first layers seen from the lumen, comprehending the whole intima layer and a section of the media layer. According to OCT performance, light is backscattered when a refractive index change appears within the sample. The received signal is then proportional to the refractive index variation. The aortic wall tissue is turbid and thick, producing numerous small inhomogeneities and attenuation along penetration, making OCT images noisy and blurry. These effects reduce the original system dynamic range to approximately 40 to 60dB.

2.4 Amorphous spots identification

For the detection of specific anomalies, it is necessary to have previous knowledge about size, shape and intensity of the region surrounding each pixel and the global intensity of the image. The procedure to detect amorphous regions is divided into different sequential steps (Fig. 1).

Fig. 1 Global procedure for the delimitation of anomalous regions in OCT images. The final purpose of delimitation is the quantification of degradation in the OCT B-scans.

Download Full Size | PDF

• Air-tissue interface segmentation: the air region between the optical probe and tissue must be removed prior to the analysis. In this work, this region is eliminated by applying a strong median filter (30 pixels vertical by 60 pixels horizontal) to the OCT image and posterior Otsu’s thresholding algorithm [16] providing the air-tissue interface. The filtered image is discarded and only the air-tissue interface is preserved. The raw image below the air-tissue interface is not affected by this filtering process.
• Hessian matrix: the Hessian matrix of an image is defined as follows for every pixel f(x,y) in the image: $H_{f (x, y)} = [\begin{matrix} \frac{\partial^{2} f (x, y)}{\partial x^{2}} & \frac{\partial^{2} f (x, y)}{\partial x \partial y} \\ \frac{\partial^{2} f (x, y)}{\partial y \partial x} & \frac{\partial^{2} f (x, y)}{\partial y^{2}} \end{matrix}]$
As described in Eq. (1), for every pixel f(x,y) in the image, the second order partial derivatives must be obtained. The chosen differentiation method is the convolution with the second order partial derivatives of a selected kernel, k(x,y). As convolution is a linear method, the procedure of kernel differentiation and following convolution with the desired image, is equivalent to applying a convolution (filtering) followed by differentiation of the filtered image resulting from the convolution. This is depicted in Eq. (2).
$\frac{\partial (f (x, y) * k (x, y))}{\partial x} = \frac{\partial f (x, y)}{\partial x} * k (x, y) = f (x, y) * \frac{\partial k (x, y)}{\partial x}$
The relation described in Eq. (2) is only true when functions f(x,y) and/or k(x,y) are differentiable, and can be applied for a second order differentiation, third order, etc as long as this condition is satisfied. The main advantage is that only the Gaussian function has to be differentiated, instead of differentiating every image.
As mentioned previously, the anomalies observed in OCT images of the aortic wall present amorphous shapes of different sizes and axial directions, being frequently horizontally elongated [2]. As there is not a relevant geometric function to model these amorphous shapes, the two dimensional Gaussian function in Eq. (3) has been assumed as the kernel for the derivation, k(x,y), also with smoothing and border enhancement purposes.
$G_{(x, y)} = \frac{1}{2 π σ_{x} σ_{y}} e^{- (\frac{{(x - μ_{x})}^{2}}{2 σ_{x}^{2}} + \frac{{(y - μ_{y})}^{2}}{2 σ_{y}^{2}})}$

The second order derivatives used for convolution are presented in Eq. (4) and its graphical representation, dependent on the value of the kernel parameters, is shown in Fig. 2.

Fig. 2 Second order partial derivatives of the Gaussian kernel for different σ_x and σ_y values.

Download Full Size | PDF

{\begin{cases} \frac{\partial^{2} G_{(x, y)}}{\partial x^{2}} \equiv G x x = (\frac{{(2 x - 2 μ_{x})}^{2}}{8 π σ_{x}^{5} σ_{y}} - \frac{1}{2 π σ_{x}^{3} σ_{y}}) e^{- (\frac{{(x - μ_{x})}^{2}}{2 σ_{x}^{2}} + \frac{{(y - μ_{y})}^{2}}{2 σ_{y}^{2}})} \\ \frac{\partial^{2} G_{(x, y)}}{\partial y^{2}} \equiv G y y = (\frac{{(2 y - 2 μ_{y})}^{2}}{8 π σ_{x} σ_{y}^{5}} - \frac{1}{2 π σ_{x} σ_{y}^{3}}) e^{- (\frac{{(x - μ_{x})}^{2}}{2 σ_{x}^{2}} + \frac{{(y - μ_{y})}^{2}}{2 σ_{y}^{2}})} \\ \frac{\partial^{2} G_{(x, y)}}{\partial x \partial y} \equiv G x y = \frac{(2 y - 2 μ_{y}) (2 y - 2 μ_{y})}{8 π σ_{y}^{3} σ_{y}^{3}} e^{- (\frac{{(x - μ_{x})}^{2}}{2 σ_{x}^{2}} + \frac{{(y - μ_{y})}^{2}}{2 σ_{y}^{2}})} \\ \frac{\partial^{2} G_{(x, y)}}{\partial y \partial x} \equiv G y x = \frac{\partial^{2} G_{(x, y)}}{\partial x \partial y} \end{cases}

being σ_x and σ_y the standard deviation in the x and y dimensions respectively and µ_x and µ_y the Gaussian mean values. The kernel size in pixels is defined as 6 times the biggest σ_x or σ_y, in order not to truncate the kernel response.

The decomposition of the Hessian matrix can be expressed according to the following expression:

H v = λ v

where the solutions are the eigenvalues λ₁ and λ₂ and the eigenvectors v₁ and v₂ that satisfy the equality condition, being λ₁<λ₂. The signs of the eigenvalues show the evolution of the backscattering intensity signal of adjacent pixels. This procedure is applied to each pixel on the image to be analyzed obtaining a pair of eigenvalues (λ₁, λ₂) and eigenvectors (v₁, v₂) for each pixel.

• Hessian analysis: the relation between λ₁ and λ₂ provides information about the shape of the region in the image [3]. For the detection of amorphous shapes, only λ₁, which represents the direction of less variation on the backscattering signal, is needed. The sign of the eigenvalues is due to the intensity of the region. By selecting only negative eigenvalues, λ₁<0, dark spots can be detected, which is the case study of this work. For every possible i combination of (σ_x, σ_y), the obtained λ₁ is normalized, obtaining the parameter A_i(σ_x, σ_y): $A_{i} (σ_{x}, σ_{y}) = {\begin{matrix} 0 & λ_{1} < 0 \\ \frac{λ_{1}}{\max (λ_{1})} & λ_{1} \geq 0 \end{matrix}$
This parameter is computed for every pixel, accounting for intensity variations in its neighborhood.
• Anomaly delineation: when the Hessian analysis is applied pixel by pixel, its response is limited to the size of the applied Gaussian kernel (defined as the 6 times the biggest σ_x or σ_y). For this reason, after the application of the Hessian analysis, Otsu’s method [16] is applied to evaluate the intensity relevance according to the global image intensity histogram. Otsu’s thresholding procedure is used to automatically delimit the anomalies according to the A_i values in Eq. (6). This procedure finds the optimum threshold to binarize the image according to the histogram distribution of the processed eigenvalue in the whole image. As a result, a mask is generated with the delimitation of intensity anomalies detected in the image.
• Accuracy estimator metric: edge detection accuracy becomes difficult to evaluate without further information about the real anomaly extension. A gold standard segmentation procedure, such as manually delineation, is not feasible in terms of time consumption, automation, and is also loose due to faint anomaly edges present in the sample set images.
A preliminary evaluation of the attained delineation accuracy is done by visual inspection, but this method presents subjectivity. To supplement this method, a new accuracy estimator metric, Δ, is proposed in Eq. (7) to evaluate the best delineation performance.
$Δ = | {\bar{I}}_{Anomalies} - {\bar{I}}_{Enclosing} |$

After the delimitation of anomalies (Fig. 3), the mean intensity of detected anomalies ( ${\bar{I}}_{Anomalies}$ , Fig. 3(b)) is computed and compared with the mean intensity of the area that surrounds and encloses the anomalies ( ${\bar{I}}_{Enclosing}$ , Fig. 3(c)). The dimension of the enclosing area is defined by a width σ_x (Fig. 3(c)).The difference of mean intensities of those areas is used as the accuracy estimator.

Fig. 3 Simulated anomaly (a), detected anomaly with border in green (b) and area enclosing the detected anomaly (c). The size of the surrounding area is defined by a width σ_x. The mean grey level intensity (8 bits, 0 to 255 levels) is 218.35 in (a), 91.2 in the delimited section in (b) and 130.0 in the delimited section in (c).

Download Full Size | PDF

When the delineation is accurate, the anomaly will be ideally dark and the surrounding ideally bright, presenting big differences in the mean intensity. Delineation errors will reduce this mean intensity difference. As real anomalies are not perfectly bright or dark, this metric is only viable for comparison of performance in the same image.

• Multiscale aggregate: as stated, the response of the differentiation and filter kernel is a function of the standard deviation (σ_y and σ_x, manually selected by the user according to an approximate expected anomaly size) and the center point (µ_x and µ_y, equal to zero for filtering purposes). The kernel width is defined by the standard deviation, which is related to the dimension of the anomaly on x and y axis. In addition to the dimension of the standard deviation, the ratio among σ_y and σ_x has been used to evaluate different shapes. $r a t i o = \frac{σ_{y}}{σ_{x}}$
In the case of ratio = 1, the spots to be detected tend to a circumferential shape. When ratio>1, σ_y>σ_x and according to Eq. (4), a horizontally elongated derivative kernel function is applied. When ratio<1, σ_y<σ_x, a vertically elongated kernel is employed.
The delineation obtained in previous steps is strongly dependent on the values selected for σ_y and σ_x and the different ratio between these two values. Following a multiscale procedure, the delineation process is repeated for different σ_y and σ_x values, obtaining different anomaly quantification, A_i(σ_x, σ_y), in each case. The final result is the one that maximizes the anomaly quantification metric, A_o:
$A_{o} = \max (A_{i} (σ_{x}, σ_{y}))$
In order to produce the best combination of different A_i(σ_x, σ_y) values, the accuracy metric Δ defined in Eq. (7) is used for automated optimization of the parameters in the multiscale aggregate. It works as follows: A_i(σ_x, σ_y) values are introduced in Eq. (9) one by one; if the addition of a new A_i(σ_x, σ_y) provides a bigger Δ accuracy metric for A_o, this A_i(σ_x, σ_y) is added to the A_o aggregate; if the Δ metric for A_o decreases, this A_i(σ_x, σ_y) is discarded.
• Degradation quantification: The area of anomalies is strictly linked to degradation in the sample [1]. A quantification score has been developed to assess degradation quantification. For each B-scan, degradation is computed as follows: $B-scan score = \sum_{k = 1}^{N a} e^{\frac{a_{k}}{512*1024}}$
where a_k is the area of every individual anomaly in pixels and Na represents the number of anomaly areas in each B-scan. The exponential provides an approximation to what is visually seen: few big anomalies are worse than many small areas in terms of degradation [1]. The number of total pixels (512*1024) in an image is used to normalize the exponent. When the degradation of the whole set of B-scans for a single specimen is aggregated, the OCT score for this specimen can be estimated:
$OCT score = \frac{1}{N b} \sum_{N b} B-scan score$
where Nb represents the number of B-scans for the sample. The term 1/Nb is included to normalize the OCT score, being independent of the number of B-scans of each specimen.

3. Results and discussion

The algorithm has been applied for the delineation on the whole set of aorta specimens described in section 2.1. For discussion purposes, only one of the worst cases of degraded aorta is displayed. Its corresponding OCT image is also affected by presence of high noise and tilt (Fig. 4) so, the validation of the method here assures its behavior under better conditions. The overall performance of the diagnostic procedure will be validated against the gold standard provided by histopathological analysis.

Fig. 4 OCT image from an aneurismatic aorta (a) and zoom for illustration purposes (b).

Download Full Size | PDF

The first steps of the analysis comprehend the air-tissue interface detection and the computation of the Hessian matrix of the image, resulting λ₁ and the consequent A_i(σ_x, σ_y) (Fig. 5). Different σ_x and σ_y will produce different kernel sizes, highlighting anomalies of different sizes and shapes.

Fig. 5 Individual A_i(σ_x, σ_y) obtained for different σ_x and σ_y parameters. Pseudo-color represents normalized A_i(σ_x, σ_y) from its minimum value (blue) to the maximum (red) with 255 colors representation. Pink lines represent the air-tissue interface computed in the pre-processing stage. Application of Otsu’s threshold method to these A_i(σ_x, σ_y) provides the anomaly delineation (black lines).

Download Full Size | PDF

In a first analysis, the dependence on the kernel size is evaluated. Different σ_x values have been applied to the selected image (Fig. 4(b)), considering an isotropic kernel σ_x = σ_y, i.e. ratio = 1. When σ_x is small, noisy areas are detected (Fig. 6(a)), whilst for large σ_x the borders of the spots are lost and delineation becomes less accurate (Fig. 6(d)). In this case, intermediate σ_x values provide a better delineation (Fig. 6(b), Fig. 6(c)). Both, visual inspection and accuracy metric Δ confirm this conclusion. The optimum result in this case, σ_x = σ_y = 7 pixels, is only valid for the OCT image under analysis. The optimum value can vary from one B-scan to another as a function of its intensity variations.

Fig. 6 Analysis of the delineation dependence on the kernel size, σ_x, maintaining an isotropic kernel with ratio = 1. Green lines represent the algorithm delineation output. Blue shaded regions represent manual reference delineation.

Download Full Size | PDF

When the ratio among σ_x and σ_y differs from 1 (Fig. 7), the response of the algorithm becomes biased towards the detection of vertical or horizontal regions. In this study, ratios below 1 remark vertical features in tissue depth (Fig. 7(a)) that, in the case of the aortic wall result in wrong delimitation. On the other hand, ratios greater than 1 (Fig. 7(c)) can improve the identification of degraded areas as they tend to be horizontally elongated because the wall degradation related to the aneurysm disease is originated by the emergence of disorders in the horizontally oriented fiber structure of the media layer [2]. The tilt of the sample surface, and consequently the anomalies orientation, does not affect the accuracy performance for ratios around 1, but for large ratios (Fig. 7(d)) it results in a bad delineation.

Fig. 7 Analysis of the delineation dependence on the kernel size orientation maintaining σ_x = 7 always fixed and varying σ_y as a function of the ratio value. Green lines represent the algorithm delineation output. Blue shaded regions represent manual reference delineation.

Download Full Size | PDF

As seen in Fig. 6 and Fig. 7, the anomaly delineation performance of the Hessian analysis depends on its initial parameters. The determination of the best σ_x and σ_y is done by visual inspection of the delineation accuracy, numerically assisted by the accuracy metric in Δ (Table 1). In this case, intensity is computed from 0 to 255 levels, so the metric is expressed in this scale.

Table 1. Accuracy metric Δ, computed for different σ_x and ratio values

View Table

According to the accuracy metric on Table 1, the best delineation, i.e. the one that achieves the biggest metric, is the case σ_x = 7 and ratio = 1 (Δ = 32.4). The worst result is obtained for σ_x = 12 and ratio = 3 (Δ = 16.9). The visual approach to the information exposed in Table 1 is shown in Fig. 8. The best simple implementation, i.e. without a multiscale approach, is shown in Fig. 8(a) and compared with the case where the accuracy metric Δ has been used for feeding the multiscale aggregate (Fig. 8(d)). When Eq. (9) is applied to generate the group for the multiscale aggregate as described in Section 2.4, the best performance is achieved for a combination of σ_x = 3, 5, 7 and ratio = 1 (Δ = 34.7), improving the case of a simple implementation σ_x = 7 and ratio = 1.

Fig. 8 Comparison of the best single application of the algorithm (top, σ_x = 7, ratio = 1) and the best multiscale aggregate according to Δ (bottom, σ_x = 3,5,7 and for each of them ratio = 1). On the first column (a, d), the original image with anomalies remarked in green and enclosing area in blue. The second row (b, e), present the segmentation of the anomalies with its mean intensity (I). The third column (c, f) represents the regions that surround the anomalies, required for the estimation of the accuracy metric, and its mean intensity (I).

Download Full Size | PDF

Manual delineation is a laborious and not obvious task because it is affected by bias and unaccuracies, even in the case of trained users, that cannot neither perceive blurred nor small intensity variations. When automated delineation (Fig. 9(c)) is compared with the manual one (Fig. 9(b)) extra anomalies appear. If both cases are compared with the original image, before delineation (Fig. 9(a)), it can be seen how manual delineation misses small dark anomalies, mainly due to the fact that they present diffuse margins.

Fig. 9 Comparison of the original image (a) with different anomaly delineation tests. In (b), the manual delineation is shown. In (c), the result of the A_o aggregate is displayed. The individual A_i(σ_x, σ_y) forming part of the final A_o are displayed from (d) to (f).

Download Full Size | PDF

The different A_i(σ_x, σ_y) included in the final A_o delineation (Fig. 9(d) to Fig. 9(f)), provide segmentation of different types of anomalies. The smallest kernel (Fig. 9(d)) delineates small inhomogeneities, being imprecise in the case of big anomalies. On the other side, big kernels (Fig. 9(f)) miss small anomalies. The aggregate combines all individual segmentations providing the best delineation according to the metric Δ. The relevance of small or big anomalies is studied later and validated against the degradation magnitude from the physiological condition of the samples.

The final delineation of anomalies in two different OCT images from two aorta specimens is shown in Fig. 10, applied in the case of severely degraded (upper row) and lowly degraded (lower row) specimens. OCT images (Fig. 10(b), Fig. 10(d)) and conventional histology images (Fig. 10(a), Fig. 10(c)) are processed for anomaly delineation. Histological images, the gold standard, evidence the real degradation of the inner layers of the aortic wall. In the case of severe degradation, anomalies in OCT images are correctly delimited (Fig. 10(a), Fig. 10(b)), presenting a wide extension. In the case of low degradation (Fig. 10(c), Fig. 10(d)), the delimited areas are smaller, and related to small inhomogeneities, in the surface (in the case of OCT) or at a cellullar level (in the case of histology).

Fig. 10 Multiscale delineation result for a severely degraded sample (top) and lowly degraded sample (bottom). Histology images (Elastic fibers-Verhoeff’s Van Gieson stain) are also processed (a, c) for comparison with OCT images of the same samples (b, d respectively). Note that dimensions and resolution are different in the case of OCT and histology. Although images are obtained in the same sample position, perfect correlation is not possible due to histology handling and cutting procedure.

Download Full Size | PDF

After the anomaly delineation, the delimited areas are processed to quantify the degradation of the vessel wall. The calculation of degradation on every B-scan according to Eq. (10), is displayed in Fig. 11 in a boxplot representation. For every specimen, each B-scan produce a slightly different B-scan score, giving a degradation fluctuation.

Fig. 11 OCT degradation score calculated on a B-scan basis. The boxplot representation includes the variation of all B-scans estimated from the same specimen. Red line: median of all the B-scans of each specimen (50% percentile). Blue box: 25 to 75% percentile. Black line: approx. 99.3% percentile. Red dots: outliers.

Download Full Size | PDF

Figure 11 shows how B-scan results differ within the same specimen providing a statistical variation and manifesting that vessel degradation is not a continuous spatial phenomena [1, 2, 13–15]. In the case of boxplots presenting a small deviation, e.g. specimen 22, there is a strong statistic, giving close degradation results for all the B-scans in the same specimen. In the case of boxplots presenting higher deviations (specimens 1, 5, 11, 15 and 21), the quantification differs strongly among B-scans suggesting strong degradation in different areas, but not along the whole specimen. When these samples are closely analyzed, it is observed that they correspond to severely degraded samples, as stated by the pathologists. The case of sample 11 is an exception; there was damage only in a portion of one of the regions of the specimen. During the histological analysis, this small region was not seen by the pathologists, and the rest of the specimen presented good health conditions.

According to (Eq. (11), the whole B-scan information is summarized into a global OCT score for each specimen. This quantification is shown in Fig. 12(b). The degradation score provided by the pathologist’s procedure is shown in Fig. 12(a) for comparison.

Fig. 12 Histological degradation quantification score (a) and OCT degradation quantification score (b).

Download Full Size | PDF

The gold standard histology score provided in Fig. 12(a), shows how the origin of samples (donor or aneurysm) does not imply null or severe degradation. The cases of sample 1, 5, 15 and 21 show the highest histology score, meaning they present the biggest degradation of all samples. Donor samples present scores from 0 to 5, as well as many of the aneurysm samples.When compared with the information obtained by the OCT score, results are comparable. Severe degradation samples present a greater OCT score than other samples, and the rest of samples provide lower and near to each other results.

In the case of histological analysis, only one histological section was processed for each specimen due to the difficult handling of excised aortic tissue. However, in the case of OCT analysis, it comprehends various B-scans (5 to 20 B-scans per specimen). As degradation is not homogenous along and aside the whole aortic vessel, the quantification of degradation will differ among B-scans [1, 2] due to this uneven spatial co-registration. However, severe degradation specimens, where patient’s risk is high, are highlighted both in the histological and OCT analysis. This makes OCT a promising technique for in-vivo evaluation of aortic vessel degradation. Tissue sampling could be even higher in OCT, reaching one B-scan each 10 µm at a rate of 1.6 B-scans per second. Therefore, the non-invasive and fast acquisition of OCT images, jointly with its high spatial resolution over the tissue surface, makes this imaging technology a promising alternative to generate degradation maps of the aortic vessel able to guide surgeons during their interventions.

4. Conclusion

The delineation of anomalous regions in OCT images of turbid media such as the aorta artery is a demanding task. In case of aneurysm disease, and to be able to guide surgeons during real-time interventions, this delineation is required to assess about the degradation status of the aorta. Diseased vessels present amorphous shapes along the wall thickness whose diffused borders pose a limitation for the delineation purposes. A solution, based on the Hessian matrix analysis, has been presented providing accurate delineation results. The proposed procedure depends on specific parameters of the Hessian analysis that have been automatically optimized using an accuracy estimator metric that avoids any heuristic approach. After the delineation, the detected degraded regions are classified according to their degradation dimensions. This method is applicable to other types of images with small modifications.

The method has been compared with histological grading techniques, providing complementary results but with higher spatial scanning resolution in the case of OCT. The quantification of degradation provided by OCT exhibits a statistical variation among regions of the same specimen that evidences the disperse spatial evolution of the physiological phenomena of vessel degradation. OCT measurements provide a different approach to the interpretation of results. Firstly, OCT can be used as a complementary method during histological analysis. Statistical variations can indicate whether there is a strong fluctuation in the sample degradation per regions, hence, a more extended histological analysis (more regions under analysis) should be performed. Secondly, OCT shows capability for detecting strong degradation produced in the vessel wall. This feature, and the fast acquisition rate, make OCT systems a feasible substitute/supplement of conventional histology techniques during surgical intervention, being transferable to multiple areas of interest.

Acknowledgments

This work has been supported by projects DA2TOI (FIS2010-19860) and FOS4 (TEC2013-47264-C2-1-R).

References and links

1. J. F. Matthias Bechtel, F. Noack, F. Sayk, A. W. Erasmi, C. Bartels, and H. H. Sievers, “Histopathological grading of ascending aortic aneurysm: comparison of patients with bicuspid versus tricuspid aortic valve,” J. Heart Valve Dis. 12(1), 54–61 (2003). [PubMed]

2. E. Real, J. F. Val-Bernal, J. M. Revuelta, A. Pontón, M. C. Díez, M. Mayorga, J. M. López-Higuera, and O. M. Conde, “Identification of vessel wall degradation in ascending thoracic aortic aneurysms with OCT,” Biomed. Opt. Express 5(11), 4089–4100 (2014). [CrossRef] [PubMed]

3. A. F. Frangi, W. J. Niessen, R. M. Hoogeveen, T. van Walsum, and M. A. Viergever, “Model-based quantitation of 3-D magnetic resonance angiographic images,” IEEE Trans. Med. Imaging 18(10), 946–956 (1999). [CrossRef] [PubMed]

4. D. J. Kroon, “Hessian based Frangi Vesselness filter,” (2010) http://www.mathworks.com/matlabcentral/fileexchange/24409-hessian-based-frangi-vesselness-filter.

5. J. Lee, J. Y. Jiang, W. Wu, F. Lesage, and D. A. Boas, “Statistical intensity variation analysis for rapid volumetric imaging of capillary network flux,” Biomed. Opt. Express 5(4), 1160–1172 (2014). [CrossRef] [PubMed]

6. Q. Li, S. Sone, and K. Doi, “Selective enhancement filters for nodules, vessels, and airway walls in two- and three-dimensional CT scans,” Med. Phys. 30(8), 2040–2051 (2003). [CrossRef] [PubMed]

7. J. Liu, J. M. White, and R. M. Summers, “Automated detection of blob structures by Hessian analysis and object scale,” IEEE 17th International Conference on Image Processing (ICIP), Hong Kong, 841–844 (2010). [CrossRef]

8. Y. Sato, S. Nakajima, N. Shiraga, H. Atsumi, S. Yoshida, T. Koller, G. Gerig, and R. Kikinis, “Three-dimensional multi-scale line filter for segmentation and visualization of curvilinear structures in medical images,” Med. Image Anal. 2(2), 143–168 (1998). [CrossRef] [PubMed]

9. S. Yousefi, J. Qin, Z. Zhi, and R. K. Wang, “Label-free optical lymphangiography: Development of an automatic segmentation method applied to optical coherence tomography to visualize lymphatic vessels using Hessian filters,” J. Biomed. Opt. 18(8), 086004 (2013). [CrossRef] [PubMed]

10. S. Yousefi, T. Liu, and R. K. Wang, “Segmentation and quantification of blood vessels for OCT-based micro-angiograms using hybrid shape/intensity compounding,” Microvasc. Res. 97, 37–46 (2015). [CrossRef] [PubMed]

11. K. M. Poole, C. A. Patil, C. E. Nelson, D. R. McCormack, M. C. Madonna, C. L. Duvall, and M. C. Skala, “Longitudinal study of arteriogenesis with swept source optical coherence tomography and hyperspectral imaging,” Proc. SPIE 8934, 89341Z (2014). [CrossRef]

12. M. Kirby, A. M. D. Lee, T. Candido, C. Macaulay, P. Lane, S. Lam, and H. O. Coxson, “Automated segmentation of porcine airway wall layers using optical coherence tomography: Comparison with manual segmentation and histology,” Proc. SPIE 8927, 8927D (2014).

13. E. Real, J. F. Val-Bernal, J. M. Revuelta, A. Pontón, M. Calvo Díez, M. Mayorga, J. M. López-Higuera, and O. M. Conde, “Enhanced delineation of degradation in aortic walls through OCT,” Proc. SPIE 9312, 931233 (2015). [CrossRef]

14. E. Real, J. F. Val-Bernal, A. Pontón, M. Calvo Díez, M. Mayorga, J. M. Revuelta, J. M. López-Higuera and O. M. Conde, “OCT for anomaly detection in aortic aneurysm resection,” IEEE Sensors Conference, Valencia, Spain, 2–5 Nov., pp. 694– 697 (2014).

15. E. Real, A. Eguizábal, A. Pontón, M. C. Díez, J. Fernando Val-Bernal, M. Mayorga, J. M. Revuelta, J. M. López-Higuera, and O. M. Conde, “Optical coherence tomography assessment of vessel wall degradation in thoracic aortic aneurysms,” J. Biomed. Opt. 18(12), 126003 (2013). [CrossRef] [PubMed]

16. N. Otsu, “A Threshold Selection Method from Gray-Level Histograms,” IEEE Sys Man Cyb 9(1), 62–66 (1979). [CrossRef]

	ratio
σ_x (pixels)	0.5	1	1.5	2	2
3	20.1	23.2	22.2	18.3	18.4
5	23.1	27.4	27.1	20.8	22.6
7	27.5	32.4	28.3	24.3	20.0
9	27.4	31.2	26.0	22.0	19.1
12	29.7	27.4	25.8	22.8	16.9

Hessian analysis for the delineation of amorphous anomalies in optical coherence tomography images of the aortic wall

Abstract

1. Introduction

2. Matherials and methods

2.1 Aorta samples and protocol

2.2. Histological analysis

2.3 Optical coherence tomography setup

2.4 Amorphous spots identification

3. Results and discussion

4. Conclusion

Acknowledgments

References and links

Cited By

Figures (12)

Tables (1)

Equations (11)

Biomedical Optics Express