Characterisation of focal liver lesions on computed tomography (CT) depends on correct interpretation of morphology and dynamic changes during bolus injection of contrast medium. The aim of this study was to develop a texture analysis concept for computer based interpretation of dynamic CT images. 148 focal liver lesions were investigated by serial CT. The study comprised 61 haemangiomas, 25 other benign lesions (FNH/adenomas) and 62 malignant lesions (primary or secondary). FNH, adenomas and malignant lesions were histologically proven. Diameter was 8-145 mm (mean 31 mm). Regions of interest were interactively defined. After extraction of characteristic textural features, a pattern classifier was trained. All CT series were evaluated using the "leaving-one-out" method. 134 of the 148 lesions were correctly classified (positive predictive value 0.9). Sensitivity for the presence of malignancy was 0.93 (80/86), specificity was 0.9 (56/62). False classification of a lesion was found to depend strongly on the quality of the examination (bolus intensity, positional change of the lesion due to respiratory movements).