Literature Review
3.5 Connection the Thesis with Previous Studies
The previous works have shown different techniques used to classify/ characterise the liver lesions from CT images. The segmentation process is considered the first step for semi-automated and fully automated CAD systems. However, all CAD systems share two main stages: the feature extraction and the classification/ characterisation stage.
In this thesis, particular attention was paid to the efficiency of the proposed algo- rithms as well as their accuracy. Regarding the segmentation process efficiency, the grey level-based approach was adopted for this task. This due to this approach has the advantages of a low computational cost, no training requirements and no user inter- action (Ruskó et al., 2009a; Gloger et al., 2010). The prior medical knowledge, pre- processing and post-processing were used to enhance the algorithm accuracy. Thus, the reasonable balance between accuracy, robustness and computational cost in frame- work design can offer a suitable solution for clinical use.
This chapter has also given an overview of different CAD systems for liver le- sion classification and characterisation. The majority of the proposed works used the texture features for lesion diagnosis, as the combination features achieved promising results in this task. In this thesis, the statistical features such as GLCM will be em- ployed, due to its considered the most popular method to drive the spatial statistical texture features, providing information about the spatial arrangement and intensities distribution in the image, and also outperforming other techniques such as wavelet fea- tures (Bayram et al.,2011). Furthermore, the intensity and shape features will be also used. This is because the malignant lesions differ from benign lesions not only in sur- face texture but also in shape, boundary and intensity (Nicolau et al.,2006;Assy et al.,
2009b;Murakami and Tsurusaki,2014a).
However, the liver lesion classification and characterisation accuracy is usually af- fected by detecting lesion appearance in CT image. These characteristics are observed differently according to the region of interest selection approaches. Lesion character- isation based on CT image methods, using existing ROI selection approaches, have a limit to represent all the lesion characteristics such as the relation between liver and lesion. Thus, the performance of the CAD systems using the current ROI selection methods was variable according to the feature extraction techniques. Hence, to over- come this limitation and obtain a better and more stable framework performance than the current methods, we proposed a multiple ROIs selection approach to well-represent the lesion characteristics.
It is crucial to emphasise that, to the best of my knowledge, all previous studies and others in literature used hand-designed features (such as texture, intensity, etc.) which are fed into a classifier (such as SVM, RF, etc.), in an attempt to classify liver lesions. However, the main limitation for using hand-designed features in lesion classification is that the diagnosis decision cannot be explained in human-level understanding, mak-
ing it less reliable for physicians. Hence, to overcome this drawback, the proposed lesion classification framework will benefit from the lesion characterisation. This will be done through utilising the high-level features to classify liver lesions. The use of high-level features provides a human-interpretable explanation of the lesion diagnostic decision to better-trusted diagnosis.
3.6
Conclusion
This chapter started with an overview of the various CAD systems based on CT image has been presented. Then, a categorised review of related literature on classification and characterisation was discussed. The review highlighted each one’s approaches and limitations and how the proposed system will address each of them. The disease diagnosed in most cases were malignant such as HCC, metastases, benign tumors such as cyst, hemangioma, which distinguished from healthy liver tissue. Texture features were used widely in classification tumor system such as COM, RLM, FOS, FM, LTE and GLCM as mentioned in the literature. Different datasets (tumors types, tumors size, dataset size, image resolution) were used in different classifiers. Furthermore, different evaluation approaches were used to calculate the system performance. As a result, it is difficult to infer the best features and classifier method which can be used in classification/ characterisation. However, some texture features have shown to be reliable in different CT image phases. For instance, GLCM and COM texture feature has been utilised successfully to enhance the classification accuracy for different CT image phases (non-enhanced and enhanced images). While FM feature was frequently considered just for one phase of CT image (non-enhanced). In addition, FOS and RLM features were most often used for enhanced CT image (after administration of contrast agent). Texture features gained more attention, while Intensity and shape features used in segmentation and lesion detection. Combination features (texture, intensity, and shape) gained promising results compared to texture features.
As overall conclusion, most of literature work was diagnosed liver lesion by ex- tracting the features from the lesion only and not paying much attention to the relation between lesion and surrounding area. The selected ROI in existing works have a limit for representing all liver lesion characteristic. Thus, the performance of these systems was variable according to used feature extraction approaches. Moreover, there are some of the semantic features such as the lesion location cannot be estimated correctly through CBIR or machine learning. The proposed system will utilise the lesion and surrounding area to capture all the characteristic of the lesion where multiple ROIs area are selected by considering the ability that each ROI represents the kind of char- acteristics and exploiting relationships between low-level and high-level features.