Validation of a new classifier for the automated analysis of the human epidermal growth factor receptor 2 (HER2) gene amplification in breast cancer specimens

Abstract

Amplification of the human epidermal growth factor receptor 2 (HER2) is a prognostic marker for poor clinical outcome and a predictive marker for therapeutic response to targeted therapies in breast cancer patients. With the introduction of anti-HER2 therapies, accurate assessment of HER2 status has become essential. Fluorescence in situ hybridization (FISH) is a widely used technique for the determination of HER2 status in breast cancer. However, the manual signal enumeration is time-consuming. Therefore, several companies like MetaSystem have developed automated image analysis software. Some of these signal enumeration software employ the so called "tile-sampling classifier", a programming algorithm through which the software quantifies fluorescent signals in images on the basis of square tiles of fixed dimensions. Considering that the size of tile does not always correspond to the size of a single tumor cell nucleus, some users argue that this analysis method might not completely reflect the biology of cells. For that reason, MetaSystems has developed a new classifier which is able to recognize nuclei within tissue sections in order to determine the HER2 amplification status on nuclei basis. We call this new programming algorithm "nuclei-sampling classifier". In this study, we evaluated the accuracy of the "nuclei-sampling classifier" in determining HER2 gene amplification by FISH in nuclei of breast cancer cells. To this aim, we randomly selected from our cohort 64 breast cancer specimens (32 nonamplified and 32 amplified) and we compared results obtained through manual scoring and through this new classifier. The new classifier automatically recognized individual nuclei. The automated analysis was followed by an optional human correction, during which the user interacted with the software in order to improve the selection of cell nuclei automatically selected. Overall concordance between manual scoring and automated nuclei-sampling analysis was 98.4% (100% for nonamplified cases and 96.9% for amplified cases). However, after human correction, concordance between the two methods was 100%. We conclude that the nuclei-based classifier is a new available tool for automated quantitative HER2 FISH signals analysis in nuclei in breast cancer specimen and it can be used for clinical purposes.

Figure 1

Image analysis of fluorescence signals. A). Automated quantitative image analysis through the tile-sampling classifier. Non-overlapping square tile were placed on DAPI-counterstain image in order to maximize the nuclear material covered. Each tile may contain a single nucleus or portions of one or more nuclei. The software effectuated the spot count in each tile. B). Quantitative image analysis through the automated nuclei-sampling classifier. The classifier automatically recognized individual nuclei in counterstain image (same image as in Figure 1A). Nuclei that were automatically recognized and considered in the analysis showed green outlines, while nuclei that were automatically recognized but not considered in the analysis had a white outline. The software effectuated spot count in each nucleus. C). After completion of the automated image analysis, the user could improve the selection of automatically selected nuclei via interaction with the software. In this particular case, some nuclei that were not yet considered during the automatic segmentation (Figure 1B) were selected by the user, some automatically selected nuclei were deleted and some overlapping nuclei were divided (yellow line between two overlapping nuclei).

Figure 2

HER2 fluorescence in situ hybridization in amplified cases. A). Amplified case showing Homogeneously Staining Regions (HSR). HSR are large clusters of HER2 fluorescence signals indicating the presence of HER2 gene amplification in tandem repeats. B). Amplified case without HSR. HER2 signals are identifiable as single spots.