RESUMO
RATIONALE AND OBJECTIVES: To determine the diagnostic performance of an artificial intelligence system for classification of focal liver lesions, in comparison to human observers. METHODS: One hundred forty-three focal hepatic lesions were evaluated with dynamic computed tomography. The study comprised 59 hemangiomas, 24 other benign lesions (focal nodular hyperplasia, adenoma), and 60 malignant liver lesions (18 primary, 42 secondary). All lesions but the hemangiomas were histologically examined by needle biopsy. For delineation of the lesion, a region of interest was defined interactively. The pattern recognition was performed in two steps with initial extraction of textural features: training of a classifier and classification of the lesions. The accuracy of classification of hepatic lesions into three groups (hemangioma, other benign processes, malignant lesions) was tested. The results were compared with those achieved by human observers using receiver operating characteristic statistical analysis. RESULTS: The accuracy (total rate of correct diagnoses) was 90.2%. False classifications were found owing to small size, weak contrast enhancement after bolus injection, respiratory movement, and atypical morphology of the lesion. The area under the receiver operating characteristic curve was not significantly different for computer and human observers. CONCLUSIONS: The system demonstrated a diagnostic accuracy comparable to human observers. Further improvement with increasing numbers of typical computed tomographic series for training of the classifier can be expected.
