AN 04

Automatic Her2 FISH scanning using tile sampling

Thomas Lörch

MetaSystems GmbH, Robert-Bosch-Str. 6, 68804 Altlussheim, Germany
eMail: tloerch@metasystems.de
WWW: www.metasystems.de

received: August 2002

Introduction

Determination of the Her2 gene amplification status is of clinical importance for the prognosis and therapy selection for breast cancer. In the conventional analysis, a tumour tissue section slide is pre-processed and hybridised using the Vysis PathVysion Her2 FISH probe kit, and 20 single tumour cells from an invasive tumour region are visually scored in the fluorescence microscope [1]. If the ratio of the overall Her2 and centromere 17 (CEP^®-17) FISH spot count exceeds 2.0, the sample is considered to be Her2 amplified. On samples with a very high cellular density (e.g. tissue sections) it may be impossible to segment a sufficient number of single nuclei for the analysis in the available slide area. Therefore we developed a simple alternative strategy, which we call „tile sampling“. Non-overlapping equi-sized square tiles are placed in the counterstain image, maximizing the total fluorescence intensity inside the tiles. Multiple channel tile sub-images are then used for the quantitative analysis and gallery display. The tile sampling method was applied to the problem of automating the Her2/CEP17 ratio analysis of breast cancer tissue sections. The frequent occurrence of HSR (homogeneously staining regions, large clusters of FISH signals) makes it impossible to use standard FISH spot counting algorithms. Therefore the samples are automatically classified as HSR and non-HSR, and depending on the outcome of this step a suitable measurement algorithm is selected. As the analysed fields can contain a mixture of tumour cells and normal cells, the automatically determined overall Her2/CEP17 ratio frequently under-estimates the real ratio value significantly. Ratio estimation methods which can improve the accuracy of the result of the automatic analysis if two cell populations are present are described in [2]. Here we employ user defined regions of interest (ROI) to exclude the majority of normal cells in the fields of view. Both approaches are described in more detail in [3].

Figure 1.
Her2/CEP17 ratio results

Materials and Methods

Training and test data sets

Two sets of breast cancer samples were used in this study: 73 samples in the training data set, and 22 samples in the independent test data set. All tissue sections were stained with the PathVysion^® Her2 FISH probe set (Vysis, Downers Grove, USA), comprising LSI^®-Her2 (orange) and CEP-17 (green), and counterstained with DAPI. The scanning was performed using a Metafer4-MetaCyte microscope scanning system.

For all samples, tumour regions suitable for analysis were interactively located using a trackball device to move the scanning stage. Then images were fully automatically captured using a 40x/0.75 dry objective, using the list of positions defined in the initial interactive step. In the Her2 and CEP17 image channels, extended focus images were computed from 9 focus plane images captured with a spacing of 0.75 µm. To provide ground-truth ratio data, all training and test samples were visually scored according to [1]. Finally, regions of interest containing at least 75% tumour cells were interactively defined.

Placing of tiles

The tile placing algorithm tries to include as much cellular material and as little empty space as possible in the tiles. The first tile is placed at the position within the image which gives the maximum total DAPI intensity in the tile.

The position of the second tile is determined using the same criterion, with the additional condition that it does not overlap with the first one. This procedure is continued until the total DAPI intensity of the next potential tile drops below a certain percentage of that of the first tile.

HSR classification

As mentioned above we are employing different spot counting algorithms for non-HSR and for HSR samples. This makes it necessary to classify the entire sample automatically before spot counting. We do this by analysing the distribution of the object areas in the Her2 channel. Objects can be single diffraction limited FISH spots or HSRs, which are significantly larger. The best discrimination of HSRs was found for the total number of objects larger than 0.7 µm², divided by the total number of all objects.

Spot counting

In non-HSR samples Her2 FISH signals are seen as diffraction limited spots with a diameter of about 3 pixels. The following tile image processing was found to give the best spot counting performance in the training data set: 1. a Gaussian smoothing filter; 2. a TopHat filter for local background correction; 3. a Laplacian sharpening filter for enhancing the spots; 4. another Gaussian smoothing filter; 5. application of the counterstain mask. After this processing the number of objects at 40% relative intensity is used as spot count.

In HSR samples all algorithms based on Her2 signal counting gave quite poor results. Therefore we decided to use a total area measurement in the orange channel and estimate the Her2 signal count from it by quadratic regression. The optimum tile image processing we found was: 1: a TopHat filter for local background correction; 2. application of the counterstain mask. The area was then measured at 18% relative intensity.

The CEP17 signals are typical centromere signals: they are larger and brighter, but have a greater tendency to be spread out or split. The optimum tile image processing in this case is: 1. a Gaussian smoothing filter; 2. a TopHat filter for local background correction; 3. a Laplacian sharpening filter for enhancing the spots; 4. application of the counterstain mask; 5. a „Spot" averaging filter with a mask area of 0.36 µm². Then the objects were counted at a relative intensity of 12%. If the distance of two signals was less than 0.5 µm, only one of them was included in the final count.

Figure 2.
Left: Part of an example Metafer4 tile gallery with unamplified signals. Right: Amplified sample with HSR.  The number displayed in red for each tile is the Her2 spot count, the number in green the CEP17 spot count, and the number in white the Her2/CEP17 ratio value.

Results and Discussion

The algorithms described above were applied to all training (73 samples) and test (22 samples) data. Of the 95 samples, 19 (20.0%) were rejected due to slide quality. Of the remaining 76 samples, 73 (96%) were correctly classified as „amplified" or „not amplified", and 3 (4%) were classified false negative. The results are shown graphically in figure 1.

Compared to manual scoring the errors of the automatic system are comparable to the discordance between two different observers. Nevertheless there are continuing efforts to reduce the errors and the sample reject rate. Of the 3 false negative classifications, 2 are very close to the threshold ratio value of 2.0. They would not cause a wrong clinical result as all ratios within an interval around 2.0 (e.g. between 1.8 and 2.2) are interactively checked before the final ratio is reported.

Figure 2 shows a part of an example Metafer4 tile gallery of an unamplified sample, Figure 3 of an amplified sample with HSR. The number displayed in red for each tile is the Her2 spot count, the number in green the CEP17 spot count, and the number in white the Her2/CEP17 ratio value.

References

PathVysion Her2 Package Insert.
Vysis, Inc., Downers Grove, USA.


J. Piper, T. Lörch, I. Poole et al. (2002)
Computing the Her2:CEP-17 ratio of tumour cells in breast cancer tissue sections by analysis of the FISH spot counts of a tiles sampling.

In: Proceedings of Quantitative Molecular Cytogenetics 2002.


T. Lörch, J. Piper, J. Tomisek (2002)
“Tile sampling“: a new method for the automated quantitative analysis of samples with high cell density and its application to Her2 scanning.

In: Proceedings of Quantitative Molecular Cytogenetics 2002.