Unsupervised detection of liver lesions in CT images

This work presents an automatic approach for liver lesions detection in CT images. In this approach, liver is first segmented using fast and reliable semi-automatic technique. After liver segmentation, lesion detection is formulated as an unsupervised segmentation approach to alleviate tedious user interaction or prior learning requirements. The Meanshift clustering technique is utilized to separate different liver tissues in each CT slice. Consequently, a rule-based system is proposed to automatically and dynamically estimate healthy and unhealthy tissues distributions, and produces initial estimation of defected tissue. Finally, the graph cuts algorithm is employed to refine the initial detection and produces the finial lesions. Validation of the proposed approach using 15 patients' CT data shows high detection rate of 93%, which makes it an efficient initial opinion in the diagnosis process.