Characterisation of Focal Liver Lesions in Multi-Phase CT Images Using Textural and Pathological Descriptors.

Characterisation of Focal Liver Lesions using imaging techniques is a low-risk approach to help physicians in diagnosis and treatment planning. State-of-the-art techniques chiefly rely on texture-based features of deep networks or conventional feature extractor techniques. However, these techniques do not sufficiently exploit the semantic attributes to improve the description of the tumour types. In this paper, we introduce pathologic descriptors to characterise the lesions in Computer Tomography images including, the quality of the tumour's boundary, the tumour's geometry, and the lesion's enhancement. We propose a blurred edge model to characterise the quality of a tumour's border. The irregularity or homogeneity of a tumour's geometry is determined by the assessment of the central part and a rim around the lesion's boundary. We compare the tumour's appearance with the intensity of normal liver tissue and the blood pools to describe its enhancement behaviour. The proposed algorithm labelled 123 Computer Tomography data consisting of five lesion types and achieved an average accuracy of 92.7%. These results prove the superiority of pathology-based descriptors compared to textural features and their future potential for the discrimination of the lesions in the brain and lung. Moreover, the features from the 'Non-Contrast' phase do not improve classification results.