Comparison of supervised and unsupervised classifications in the detection of hepatic metastases

Automatic detection of liver metastases is one of the approaches to the diagnosis support systems. It facilitates the location of the tumor for a possible surgery and makes a good treatment strategy. However, the hypodense character of the lesions and the low contrast in CT make the diagnostic difficult. Therefore, biopsy remains the best standard method to assess liver disease. To achieve a good segmentation for liver cancer, we propose in this work an automatic system containing two different methods for detecting these lesions. The first approach is based on textural analysis upon wavelet transformation and co-occurrence matrix with a supervised learning: neural network. The second approach is based on unsupervised training technique through K-nearest neighbors. This work is a comparative study of the developed methods to determine the most appropriate of them for the detection of liver metastases. We evaluated our approach on a database comprising CT images from 11 patients. We concluded that the textural approach had the best performance with a recognition rate equal to 80.667%.