After surveying the published papers in literature, it is observed that many researchers try to diagnose liver diseases using different techniques to increase the classification performance. However, it has been found that the previous studies on CAD systems usually used the absolute value of features, which are extracted from lesion regions. As a consequence, the performance is varied significantly under different acquisition conditions. For example, the CT machines or operators are different. In this section, the surrounding normal tissue of liver in the same image is used as reference. So for a certain feature, we calculate the difference of features between the lesion and surrounding normal liver tissue and employ it as a new feature vector in our proposed classification system based on low-level features.
Feature extraction is a crucial stage in the CAD system. Understanding the corre- lation between the lesion characteristics and corresponding imaging features is critical for image training, as well as for features extraction. Traditional CAD systems usu- ally use the absolute features of the lesion area for the classification task. The major drawback of absolute features, in general, is the range of distinctive features of the same lesion type may vary, depending on the ratio of the contrast agent absorption re- sulting the amount of contrast agent injection and imaging time. Hence, to overcome this limitation, we proposed the DoF to calculate the contrasting (difference/relative) features between the lesion and surrounding area. This is due to the conspicuity of a liver lesion depends on the attenuation difference between the lesion and the normal liver. Contrast agents are usually needed to contrast the lesion and surrounding normal tissue. However, the contrast agent will be absorbed by the lesion, as well as by the liver parenchyma, but at a different ratio depending on the type of the lesion and the contrast-imaging phase. As a result, DoF technique uses contrast as a discriminative feature for lesions classification. Hence, the extracted lesion features are normalised by surrounding liver features based on the ratio of contrast agent absorption.
First of all, the proposed system defines two types of ROIs for extracting the fea- tures relating to intensity and texture. The first ROI is the lesion boundary (Rl), and
the second ROI is the surrounding normal liver tissue (Rs) as shown in Figure5.3. In
contrast with existing works about the identification of lesions using one ROI (lesion area only), we also consider the second ROI (Rs) which surrounds the first ROI (Rl).
Moreover, the second ROI will be used as well to extract low-level features. The differ- ence of features between the first ROI and the second ROI will be employed as a new feature vector. However, there are some constrains to identify the second ROI: (1) The second ROI must be centrally surrounding the first ROI. (2) The ratio of the diameter between the (Rl) and (Rs) is heuristically chosen through exhaustive experiments to be
1:1.5, as presented in Table5.1. (3) The first ROI is excluded from the second ROI region. (4) The liver vessels (V) are excluded from the (Rs) region. As displayed in
Chapter 5. Liver Lesion Characterisation and Classification
Figure 5.3: Proposed framework for Lesion and normal liver tissue ROI selection; (a) Original CT image; (b) Liver lesion and vessels detection; (c) First ROI is red border for segmented lesion and second ROI is green border for surrounding normal liver tissue; (d) Actual segmented lesion; (e) Normal liver tissue excluding lesion area and liver vessels.
Table 5.1 depicts the experiments results of DoF technique using different ratio between lesion and surrounding liver on the classification accuracy over our Dataset II. Sensitivity Specificity PPV NPV ratio (Lesion:Liver) B M B M B M B M Average Accuracy 1:0 0.863 0.745 0.745 0.863 0.829 0.792 0.792 0.829 0.815 1:0.5 0.890 0.784 0.784 0.890 0.855 0.833 0.833 0.855 0.847 1:1 0.904 0.784 0.784 0.904 0.857 0.851 0.851 0.857 0.855 1:1.5 0.945 0.863 0.863 0.945 0.908 0.917 0.917 0.908 0.911 1:2 0.932 0.843 0.843 0.932 0.895 0.896 0.896 0.895 0.895 1:2.5 0.880 0.765 0.765 0.880 0.859 0.796 0.796 0.859 0.836
Table 5.1: The experiments results of using different ratio between lesion and sur- rounding liver on the classification accuracy.
Table 5.2presents the algorithm for extracting the normal liver tissue surrounding the lesion. Where the segmented area will be assigned as a second ROI.
ALGORITHM 1: Segmenting Surrounding Normal Liver Tissue (Rs)
INPUT:
Segmented Liver (LV); Segmented Liver Vessels (V); Segmented Lesion (Rl);
OPERATION:
FIND MAXDIM= maximum diameter of Rl;
CREATE Structure element, X_OU T SIDE = 1.5 × M AXDIM ; SET ¯R = Dilate(Rl, X_OU T SIDE) ∈ LV ;
SET Rsv = ¯R − Rl;
SET Rs = Rsv− V ∈ Rsv;
OUTPUT (Rs);
Table 5.2: The proposed programming algorithm to extract the second ROI from nor- mal liver tissue that surround the lesion.
The Figure 5.4depicts proposed framework to classify liver lesion based on low- level features by applying our proposed difference-of-feature technique.
Figure 5.4: Proposed framework for Lesion classification based on difference-of- feature technique.
Chapter 5. Liver Lesion Characterisation and Classification
In our system, the low-level features, as discussed in Section 5.2.1, are extracted from the lesion and surrounding normal liver tissue area, and used the difference be- tween them to generate a new feature vector and used in training a classifier. Equation
5.17depicts example to calculate mean (µ) by applying difference-of-feature. dif f erence(µ) = 1 N X (x,y)∈Rs IN ormal(x, y) − 1 M X (x,y)∈Rl ILesion(x, y) (5.17)
Where IN ormal(x, y) means the gray level at pixel (x,y) of normal surrounding liver
tissue Rs, ILesion(x, y) means the gray level at pixel (x,y) of lesion Rl, M is the total
number of pixels inside the Rs of normal liver and N is the total number of pixels
inside the Rlof lesion.