CAPÍTULO 3: Las Estrategias de Ingreso en las Facultades de Ciencias Agrarias y
3. La propuesta de cambio en la Estrategia de Ingreso de la Facultad de Ciencias
In contrast with the previous techniques of using texture only to classify liver lesions, a mix of texture with other features (intensity, shape), were used in the literature as well. Table3.4 displays a different classification system based on combined features (Intensity, texture and shape features).
In 2005, (Lee et al., 2005) proposed a system to classify three different types of lesions, namely: Cyst, Hemangioma, and HCC. Lesions were drawn by experienced radiologist from non-contrast CT image. The combination features (mean gray level, entropy, local variance) were extracted to fed Back-Propagation Cerebellar Model Ar- ticulation Controller (BP-CMAC) neural network classifier. The accuracy of proposed system recorded 87%. However, the proposed system based on extracted features di- rectly from given ROI not provide robust and accurate performance.
Later, the approach was extended by utilising shape descriptor in addition to GLCM features (Lee et al.,2007). A sequential forward selection algorithm was used to reduce feature space and adopted 4-layer pyramid scheme in classification. First layer in clas- sifier distinguished between normal and abnormal liver tissue and cyst from abnormal liver tissue in the second layer. The third layer used to identify Hemangioma and last
layer recognised HCC from undefined liver tissues. Three different types of classifier were used, namely: SVM, MILP, and RBF neural network. The accuracy of classifi- cation was 89.5%, 82.1%, and 86.7% respectively. However, the proposed system is confused with hepatoma and cavernous hemangioma diseases and cause false positive results. Recently, the same approach was improved by utilising four features from the ROI, namely: edge, roundness, contrast, and internal texture (Lee et al., 2014). The extracted feature sets were fed to SVM classifier to classify the lesions with accuracy 93.7%.
Another approach based on portal phase of contract enhancement CT image was proposed by (Yang et al., 2013b). The proposed system was able to detect and clas- sify three types of tumors, namely: hyperdense, hypodense, and heterogeneous. The combination features (Fast discrete curvelet transform, Biorthogonal wavelet, His- togram, and intensity superpixel) were extracted to fed Naive Bayes Nearest Neighbor (NBNN). The accuracy for classification has achieved 93%. However, the important advantage of nonparametric method when compared with other methods, that required training for detection of liver lesions and classification and need adjust the algorithm parameters carefully, which makes it flexible and easy to implement. On the other hand, Image dataset and the local descriptors are considered as main role in the detec- tion performance and only depend on single-phase CT slices to detect lesion.
Additional study in liver lesion classification was provided by (Doron et al.,2014b). The combination of texture features (GLCM, LBP, Gabor, GLBP) and intensity fea- ture (gray level intensity) are obtained from a given lesion. For classification module, SVM and KNN classifier were used to distinguish between four types of liver tissues, namely: Cyst, Hemangioma, Metastases, and Healthy tissue. The best result of 97% accuracy was obtained with combination of Gabor, LBP and Intensity features using SVM classifier. However, the main disadvantage of LBP feature is that the spatial rela- tions among LBPs are often eliminated within the LPB histogram generation process, because they are selected in a single histogram and results to loss of global image in- formation (Guo et al., 2010). While, computation complexity is considered the main main disadvantage of Gabor wavelet, due to producing a large number of redundant features at different scales (Baaziz et al.,2010).
The semi-automatic lesion classification system based on multiphase CT images was proposed by (Chang et al.,2017). Three types of features were extracted from the lesions in each phase, including texture, shape and kinetic curve. The GLCM texture feature was calculated 3D texture data of the lesion. The 3D shape features were obtained using compactness, elliptic model and margin to describe the lesion shape. A kinetic curve was created from each phase of CT image to represent differences in density of the lesion between each phase. The most useful features were found using a feature selection, based on Backward elimination. The extracted feature sets were fed to binary logistic regression classifier to classify the lesions with leave-one-out
Chapter 3. Literature Review
cross validation. A total of 71 cases including 49 benign and 22 cases of malignant were used to evaluate classification performance. The highest accuracy of 81.69% was achieved through combining all of the features. However, the majority of the dataset cases were benign lesions, where only 30% of the cases are malignant.
Author Year Dataset Features Accuracy
(Lee et al.,2005) 2005
Cyst (55); Hepatoma (33);
Hemangioma (33)
Mean gray level; Entropy; Local variance 87% (Lee et al.,2007) 2007 Cyst (76); HCC (30); Haemangiomas (40) GLCM; Shape descriptor 89.5% (Yang et al., 2013b) 2013 Hyperdense; Hypodense; Heterogeneous Fast discrete curvelet transform; Biorthogonal Wavelet; Histogram; Intensity superpixel 93% (Lee et al.,2014) 2014 Cyst (76); HCC (30); Haemangiomas (40) Edge; Roundness; Contrast; Internal texture 93.7% (Doron et al., 2014b) 2014 Cyst (43); Haemangiomas (24); Metastases (25); Healthy tissue (20) GLCM; LBP; Gabor; GLBP; Gray level intensity 97% (Chang et al., 2017) 2017 Malignant (22); Benign (49) 3D texture (GLCM); 3D shape (Compactness, Elliptic model and Margin); Kinetic curve 82.69%
Table 3.4: Liver lesion classification literature work based on combined features (in- tensity, texture and shape features).
In summary: According to the previous literature about liver lesion classification based on texture and combination features extracted from lesion ROI, the majority of researcher are focused on feature extraction, and usually using the absolute value of features that extracted from lesion area. Moreover, the statistical texture features gained more attention comparing to other types of texture features or intensity and shape feature (Duda et al., 2004). However, the characteristics of malignant lesions
differ from benign lesions in terms of shape, boundary and effect on surrounding liver tissues (Nicolau et al., 2006; Assy et al., 2009b; Murakami and Tsurusaki, 2014a). Furthermore, the classification through black box low-level features is meaningless for radiologist because it does not provide the understandable information behind the classification decision. There exists a limited work that benefited from all the char- acteristics of the lesion such as border, shape and surrounding area. In addition, in- terpreting the classification results through high-level features. Hence, our proposed system utilises all the lesion characteristics (internal, border and surrounding area) with the combined between intensity, texture and shape features to enhance the system accuracy. In addition, proposing multiple ROIs to calculate the high-level features by considering the ability of each ROI that represents a set of characteristics, and then using high-level features to classify the lesion. In contrast with most existing research, which use low-level features only, the use of high-level features and characterisation helps in interpreting and explaining the classification and is more intuitive to clinicians.