Comparison of pathologist-detected and automated computer-assisted image analysis detected sentinel lymph node micrometastases in breast cancer.

Sentinel lymph node biopsy has stimulated interest in identification of micrometastatic disease in lymph nodes, but identifying small clusters of tumor cells or single tumor cells in lymph nodes can be tedious and inaccurate. The optimal method of detecting micrometastases in sentinel nodes has not been established. Detection is dependent on node sectioning strategy and the ability to locate and confirm tumor cells on histologic sections. Immunohistochemical techniques have greatly enhanced detection in histologic sections; however, comparison of detection methodology has not been undertaken. Automated computer-assisted detection of candidate tumor cells may have the potential to significantly assist the pathologist. This study compares computer-assisted micrometastasis detection with routine detection by a pathologist. Cytokeratin-stained sentinel lymph node sections from 100 patients at the University of Vermont were evaluated by automated computer-assisted cell detection. Based on original routine light microscopy screening, 20 cases that were positive and 80 cases that were negative for micrometastases were selected. One-level (43 cases) or two-level (54 cases) cytokeratin-stained sections were examined per lymph node block. All 100 patients had previously been classified as node negative by using routine hematoxylin and eosin stained sections. Technical staining problems precluded computer-assisted cell detection scanning in three cases. Computer-assisted cell detection detected 19 of 20 (95.0%; 95% confidence interval, 75-100%) cases positive by routine light microscopy. Micrometastases missed by computer-assisted cell detection were caused by cells outside the instrument's scanning region. Computer-assisted cell detection detected additional micrometastases, undetected by light microscopy, in 8 of 77 (10.4%; 95% confidence interval, 5-20%) cases. The computer-assisted cell detection-positive, light microscopy-missed detection rate was similar for cases with one (3 of 30; 10.0%) or two (5 of 47; 10.6%) cytokeratin sections. Metastases detected by routine light microscopy tended to be larger (0.01-0.50 mm) than did metastases detected only by computer-assisted cell detection (0.01-0.03 mm). In a selected series of patients, automated computer-assisted cell detection identified more micrometastases than were identified by routine light microscopy screening of cytokeratin-stained sections. Computer-assisted detection of events that are limited in number or size may be more reliable than detection by a pathologist using routine light microscopy. Factors such as human fatigue, incomplete section screening, and variable staining contribute to missing metastases by routine light microscopy screening. Metastases identified exclusively by computer-assisted cell detection tend to be extremely small, and the clinical significance of their identification is currently unknown.