Feature extraction

denote the entirety ALSR with the index ofi, whereiis the ALSRs number depending on the value oflength_LB/l_w, wherelength_LBis the length of the vessel lumen boundary. With the multi-layer model definition, a PLBR in an individual IVOCT image can be divided into severalALSR¹²³s in the circumferential dimension to characterize the circumferential continuity and difference of tissues, and eachALSR¹²³scomposed of three ALSRs to describe the hierarchical properties of tissues. Figure 4.2 illustrates the schematic of ALSR applying to PLBR and the hierarchical structures for ALSRs with the defined parametersl_w,l_h¹,l_h²and l_h³. In our model, I define three ALSR with three different sizes ofl_hⁱ by according to the property that the different tissues present different thicknesses in PLBR. (Tab. 4.1).

Fig. 4.2 The PLBR is divided into 3 layers along the radial direction by considering the light attenuation property of different vessel tissues. A individualALSR^mis defined in each layer for the 2-D local features extraction.

4.3 Feature extraction

Through the definition of the local multi-layer model, the following characteristic descriptors are investigated and applied to extract features from an ALSR of each layer of PLBR. We move counterclockwise the ALSR with a specific stride in the circumferential dimension to obtain a local A-line cluster, and then extract features based on this local multi-layer. Let ALSR^m_i denotes thei-th ALSR of them-th layer.

Kullback-Leibler Divergence

Formally, the Kullback-Leibler divergence (KL divergence)[44] is a measure of the asymme-try of the difference between two probability distributions P and Q. For discrete probability

distributionPandQ, its KL divergence is defined as D_KL(P∥Q) =−

∑

i

P(i)lnQ(i)

P(i), (4.1)

or

D_KL(P∥Q) =

∑

i

P(i)lnP(i)

Q(i), (4.2)

where the KL divergence is valid only if for allx,Q(x)>0 andP(x)>0, simultaneously the sum of probabilityPandQrespectively equal to 1.

Here, we use the histogram of an ALSR to denote its intensity probability distribution.

The similarity of two regions (A and B, which are selected fromALSRsbelong to oneALSR^m_i ) is calculated with the symmetric form of the KL divergence presented as follows:

D_s= D_KL(P_A,P_B) +D_KL(P_B,P_A)

2 (4.3)

whereD_KL(P_A,P_B) =∑P_Alog(P_A/P_B),P_A andP_Bboth are the discrete probability distribu-tions ofAandB. Therefore, KL divergences are calculated among the three ALSRs (ASLR¹_i, ASLR²_i andASLR³_i) which have the same polar angle along the radial direction. To measure the relative attribute of any pairwise from theASLR¹²³_i , three KL values [D¹²_s ,D¹³_s ,D²³_s ] are gained through pairwise calculation.

Radial direction intensity difference

The intensities among ALSRs corresponding to regions of vessel lesions tissues show obvious intensity difference in the radial dimension because of the variance of light attenuation of each tissue Visible and valid region with a certain thickness demonstrate the above characterization in Fig. 4.1 (B). I construct an assessment factor to statistically describe the radial dimension intensity difference. This measurement expresses the sum of intensity difference of ALSRs in the same polar angle direction. It is defined as:

RDID_i=

M

∑

m=1

|e^mp_i −e^mq_i |

e^mp_i +e^mq_i (4.4)

wheree^mp_i =∥V_i^m−µ_i^p∥,e^mq_i =∥V_i^m−µ_i^q∥,V_i^mis the 2-D pixel intensity matrix ofALSR^m_i that is thei-thALSR in them-thlayer,µ_i^pandµ_i^qrespectively indicate the average intensity ofALSR_i^pandALSR^q_i. Them,p,q∈ {1,2, . . . ,M}are used to denote the layer index ,M is the maximum layer number andm̸= p̸=q. Radial direction intensity difference (RDID)

4.3 Feature extraction 73 can measure the statistic information of the intensity variance of ALSRs in the same angle direction to characterize the local adjacent A-lines intensity alteration.

Accumulated circumference difference

Observing from Fig. 4.1, at the same layer, ALSRs with the same tissue class have similar features and appearance while the texture characteristics between different tissues exist in obvious contrast. That is, if measuring the distribution of vascular tissues in the circumferen-tial domain, the angle range used for the measurement of tissues is a significant quantitative characterization. In a homogeneous tissue, the intensity variation circumferentially of ALSRs presents less difference, but for the different tissues, the variance of these ALSRs appear outstanding. To express this statistical character, an accumulated intensity difference of ALSR^m_i in themth layer is designed:

ACD^m_i =

K−1

∑

|µ_i^m−µ_i,^m_j|

σ_i,^m_j (4.5)

whereµ_i^mandµ_i,j^m are the average intensity of theALSR^m_i andALSR^m_i,jrespectively. ALSR^m_i,j is a joint region composing of ALSR^m_i and ALSR^m_j in themth layer. σ_i,^m_j is the standard deviation ofALSR^m_i,j. ParameterKis the total number of ALSRs of a single layer.

Depth of region of interest (ROI)

The aforementioned content[11, 18, 28, 86, 88, 91, 96] discussed that lipid, fibrous and calcified plaques contain different penetration depths and attenuation coefficients when utilizing the catheter to capture the vessel inner structure situation. Therefore, the ROI defined in an IVOCT image should contain almost all useful information about the vessel tissues. In this paper, we applied the Chan-Vese-level-set method[20] to obtain an energy dividing line as the outer boundary of the ROI. Corresponding to the number of the ALSRs of a single layer, ROI is split intoKparts, computing the average distance between the points of lumen boundary and the corresponding points on the outer boundary for each ALSR part as the distance feature of the ROI. The set of distanceDis defined:

D_ROI={d_k|k=1,2, . . . ,K} (4.6) whered_kis the average distance ofkth part of the ROI.

Gray level co-occurrence matrix

The texture filter functions provide a statistical view of texture based on the image histogram.

These functions can provide useful information about the texture of an image but cannot provide information about shape, i.e., the spatial relationships of pixels in an image. Gray level co-occurrence matrix (GLCM)[31] calculates how often a pixel with the gray value ioccurs in a specific spatial relationship (based on the angleθ and distanced) to a pixel with the value j. To am×nimageIthat its gray level is supposed asG(the gray range of [1,G]), letCbe a matrixN×N, andC_∆x,∆y(i, j)denotes that the number of times inIof the adjacent pixelsiand j, the adjacent relation is given by∆xand∆y,C_∆x,∆y(i, j)is defined as the following formula:

C_∆x,∆y(i, j) =

n p=1

∑

m q=1

∑







1, i f I(p,q) =i and I(p+∆x,q+∆y) = j, 0, otherwise.

(4.7)

For an individual ALSR unit, statistical measures of GLCM could be calculated as the features of ALSR. Here, we utilized 4 statistical features considered for the analysis of image texture information. The properties are computed as follows:

Dissimilarity:

fdissimilarity=

N−1 i,

∑

C_i,j|i−j| (4.8)

Homogeneity:

fhomogeneity=

N−1 i,j=0

∑

C_i,j

1+ (i−j)² (4.9)

Energy:

f_energy=

N−1

∑

i,j=0

C_i,j² (4.10)

Correlation:

fcorrelation=

N−1 i,j=0

∑

C_i,j





(i−µi)(j−µj) q

σ_i²σ²_j



 (4.11)

whereC_i,jis the co-occurrence matrix,iand jare the labels of the columns and rows of the GLCM.µ is the mean andσ is the standard deviation. In this paper, we tested and chose the distanced=3 and angleθ = [0, 90, 180, 270 deg] of GLCM, 16 texture features for each ALSR^m_i are selected as its features.