Objectives: As part a Program Project to evaluate emerging optical technologies for cervical neoplasia, our group is performing quantitative histopathological analysis of biopsies from 1800 patients. Several methodological issues have arisen with respect to this analysis: (1) Finding the most efficient way to compensate for staining intensity variation with out losing diagnostic information; (2) Assessing the inter‐ and intra‐observer variability of the semi‐interactive data collection; and (3) the use of non‐overlapping cells from the intermediate layer only. Methods: Non‐overlapping quantitatively stained nuclei were selected from 280 samples with histopathological characteristics of normal (199), koilocytosis (37), CIN 1 (18), CIN 2 (10) and CIN 3 (16). Linear discriminant analysis was used to assess the diagnostic information in three different feature sets to evaluate and compare staining intensity normalization methods. Selected feature values and summary scores were used to evaluate intra‐ and inter‐observer variability. Results: The features normalized by the internal subset of the imaged cells had the same discriminatory power as those normalized by the control cells and by both normalization methods seem to have additional discriminatory power over the set of features which do not require normalization. The use of the internal subset decreased the image acquisition time by ∼50% at each center, respectively. The intra‐ and inter‐observer variability was of a similar size. Good performance was obtained by measuring the intermediate layer only. Conclusion: The use of intensity normalization from a subset of the imaged non‐overlapping intermediate layer cells works as well as or better than any of the other methods tested and provides a significant timesaving. Our intra‐ and inter‐observer variability do not seem to affect the diagnostic power of the data. Although this must be tested in a larger data set, the use of intermediate layer cells only may be acceptable when using quantitative histopathology.
Measurement of Face Detection Accuracy Using Intensity Normalization Method and Homomorphic Filtering