Automatic detection of liver lesion from 3D computed tomography images

Automatic lesion detection is important for cancer examination and treatment, whereas it remains challenging due to the varied shape, size, and contextual anatomy of the diseased masses. In this paper, we present a robust and effective learning based method for automatic detection of liver lesions from computed tomography data. The contributions of this paper are the following. First, we develop a cascade learning approach to lesion detection comprising multiple detectors in the spirit of marginal space learning. Second, a gradient based locally adaptive segmentation method is proposed for solid liver lesions. The segmentation results are used to extract informative features for classification of generated candidates. Extensive experimental validation is carried out on 660 volumes with 1,302 hypodense lesions, and 234 volumes with 328 hyperdense lesions, with a resulting 90% detection rate at 1.01 false positives per volume for hypodense lesion and 1.58 false positives per volume for hyperdense lesion, respectively.