Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2011 Jun 1.
Published in final edited form as: Technol Cancer Res Treat. 2010 Jun;9(3):231–242. doi: 10.1177/153303461000900302

Automated Analysis of Fluorescent in situ Hybridization (FISH) Labeled Genetic Biomarkers in Assisting Cervical Cancer Diagnosis

Xingwei Wang 1, Bin Zheng 1, Roy R Zhang 2, Shibo Li 3, Xiaodong Chen 4, John J Mulvihill 3, Xianglan Lu 3, Hui Pang 3, Hong Liu 4,*
PMCID: PMC2916642  NIHMSID: NIHMS220245  PMID: 20441233

Abstract

The numerical and/or structural deviation of some chromosomes (i.e., monosomy and polysomy of chromosomes 3 and X) are routinely used as positive genetic biomarkers to diagnose cervical cancer and predict the disease progression. Among the available diagnostic methods to analyze the aneusomy of chromosomes 3 and X, fluorescence in situ hybridization (FISH) technology has demonstrated significant advantages in assisting clinicians to more accurately detect and diagnose cervical carcinoma at an early stage, in particular for the women at a high risk for progression of low-grade and high-grade squamous intra-epithelium lesions (LSIL and HSIL). In order to increase the diagnostic accuracy, consistency, and efficiency from that of manual FISH analysis, this study aims to develop and test an automated FISH analysis method that includes a two-stage scheme. In the first stage, an interactive multiple-threshold algorithm is utilized to segment potential interphase nuclei candidates distributed in different intensity levels and a rule-based classifier is implemented to identify analyzable interphase cells. In the second stage, FISH labeled biomarker spots of chromosomes 3 and X are segmented by a top-hat transform. The independent FISH spots are then detected by a knowledge-based classifier, which enables recognition of the splitting and stringy FISH signals. Finally, the ratio of abnormal interphase cells with numerical changes of chromosomes 3 and X is calculated to detect positive cases. The experimental results of four test cases showed high agreement of FISH analysis results between the automated scheme and the cytogeneticist’s analysis including 92.7% to 98.7% agreement in cell segmentation and 4.4% to 11.0% difference in cell classification. This preliminary study demonstrates that the feasibility of potentially applying the automatic FISH analysis method to expedite the screening and detecting cervical cancer at an early stage.

Keywords: Fluorescence in situ hybridization (FISH), Automated FISH analysis, Cervical cancer, Computer-aided detection (CAD), Chromosomes 3 and X

Introduction

Cervical cancer represents approximately 9% of all cancer deaths in women, making it one of the most common causes of death in women (1). Detecting cervical cancer at an early stage is important to successfully cure more patients with effective treatments. The clinical data have supported that using the Pap-smear test, based solely on the morphologic examination of exfoliated cells from the cervix, has substantially reduced the death rate due to cervical cancer in the women (2). The development of cervical cancer is often preceded by distinct morphological changes from normal epithelium to carcinoma through low-grade and high-grade squamous intra-epithelium lesions (LSIL and HSIL). However, detecting the low-grade squamous intra-epithelium lesions and non-papillary tumors is quite difficult because these tumors shed diagnostic cells intermittently and the cells may be markedly degenerated (3). As a result, by the time a diagnosis is reached, the neoplasm is often quite large and in many instances metastasizes to regional lymph nodes. In the last decade, increased knowledge of cancer genetics has led to the development of new assays for cancer detection. One such assay, fluorescence in situ hybridization (FISH), has become a valuable molecular imaging tool to detect cervical cancer, predict its prognosis, and monitor the therapy response. Specifically, the genetic analyses and studies have validated that the status of chromosomes 3 and X were two useful intermediate biomarkers to screen and diagnose cervical carcinoma by identifying squamous intra-epithelial lesions at a high risk of progression (4).

FISH technology involves the precise annealing of a single-stranded fluorescently labeled deoxyribonucleic acid (DNA) probe to complementary target sequences and allows the visualization of specific nucleic acid sequences within a cellular preparation. After identifying that numerical changes of particular chromosomes (e.g. chromosomes 3 and X) had a significant impact on the cervical cancer development and prognosis, a number of research groups have demonstrated that compared to other available methods, FISH analysis of Pap-smear slides could achieve a higher sensitivity and specificity in detecting early cervical cancers by identifying specific changes of biomarkers (4–7). For example, Heselmeyer et. al. (7) reported that FISH sensitivity to predict progression of cytologically low-grade or normal Pap-smears to invasive carcinoma was 100% with specificity of 70%. Meanwhile, FISH analysis could also predict the progression of uterine cervical dysplasia to invasive cancer with a higher accuracy (8–10). In the cytogenetic laboratories, interphase FISH image analysis is commonly used to detect or identify the deletion/duplication and specific translocation of small parts of genes. For this purpose, the sample DNA of interphase nuclei are first denatured, which separates the complimentary strands within the DNA double helix structure. Then, the fluorescence probe is bound to the denatured sample mixture and hybridizes with the sample DNA at the target site as it backs into a double helix.

Currently, manual and semi–automated FISH analyses are routinely performed in the cytogenetic laboratory. Using a fluorescence microscope, a lab technician first visually screens and selects around 50 ~ 100 cells in each examination to minimize the FISH signal counting bias and he/she then computes a summary index. However, this manual FISH analysis method is time-consuming and labor-intensive, and also introduces inter-reader variability. As a result, development of automated FISH analysis systems has been attracting great research interests in the last several years. Netten et al. reported the first automatic system for FISH analysis (11). It reported that 89% of interphase nuclei were correctly segmented; the relative intensity and area were computed and used as two features to analyze FISH signals. It was only applied to count FISH signals that are neither split nor stringy in a single spectrum. Since in the clinical environment, various combinations of probes and FISH signals distributed in different spectrums are commonly used in the FISH analysis, more robust automated schemes with high performance should have the ability to compensate many practical issues related to the overlapped interphase cells, background noise, pseudo-artificial fluorescence noise, and fluorescence signals distributed in different focus levels. Thus, other automated FISH image analysis systems and schemes have been recently developed and tested (11–15). In these schemes, user-defined thresholds (12), neural networks (13, 14), the watershed algorithm (15), and the Isodata algorithm (11) were several popular algorithms used to segment interphase nuclei. A top-hat transform was employed to segment FISH signals (11) and component labeling algorithms (16) were implemented to count the number of FISH signals in the binary FISH images. In addition, morphological features and distances are key factors used to remove false-positive signals (primarily due to the image noises) and detect true-positive FISH signals (17). For example, Gué et al. (18) utilized the size (volume) and distances to identify FISH signals. In order to merge splitting FISH signals, different cut-off thresholds (0.5 μm (17, 19) and 0.8 μm (20)) were typically applied in these schemes.

Despite of study progress and reported encouraging results, none of the automated FISH image analysis methods have been routinely used in cytogenetic laboratories. The following factors limit the clinical utility of these automated FISH image analysis systems including: (a) relatively lower sensitivity in detecting and segmenting analyzable interphase nuclei (i.e., approximately 69%~89% (20, 21)); (b) higher false-positive detection rates (i.e., 7% ~ 18.8% reported by Kajtár (20) and Vrolijk (12)); and (c) the lack of commonly accepted standards to distinguish and merge splitting and stringy FISH signals. To overcome these limitations, we conducted this preliminary study to develop and test a new automated FISH image analysis scheme for detecting abnormal changes of chromosomes 3 and X, which aims to eventually expedite the screening of cervical cancer at an early stage and improve detection efficiency and accuracy.

Materials and Methods

A testing FISH image dataset

In this study, we randomly selected four Pap-smear specimens that were acquired from four patients who underwent annual cervical cancer screening examinations at the University of Oklahoma Health Sciences Center. Based on the established protocol in our cytogenetic laboratory, we then took the following four steps to produce FISH labeled biomarker slides from the originally acquired specimen of Pap-smear examinations:

  1. Pretreatment the pap-smear slides: The slides are first pretreated in 2 × saline-sodium citrate buffer (SSC) for 30 minutes and 0.01% pepsin/0.01 M HCI at 37°C for 13 minutes at 37°C. Second, the slides are washed in phosphate-buffered saline (PBS), post-fix, PBS respectively for 5 minutes. Finally, the slides are quenched in 70%, 85% and 100% ethanol, respectively for 1 minute followed by air drying in the hood for 10~15 minutes.

  2. Denature: The slides are immersed into the denature solution and the probe is also denatured in a water bath at 72°C for 5 minutes.

  3. Hybridization: Two centromeric enumeration probes CEP3 (D3Z1) and CEPX (DXZ1) (Vysis, Abbott Molecular Inc., Downers Grove, IL) are mixed together and applied to the processed Pap-smear slides. These mixed probes can process the Pap-smear specimens and mark chromosomes 3 and X located inside the interphase nuclei. The hybridization procedure is followed by incubation in a 37 °C moist chamber overnight.

  4. Post-hybridization: the slides are washed in post-hybridization buffer 0.3%NP-40/2 X SSC at 72 °C for 1 minute and 0.1%NP-40/2 X SSC at room temperature for 1 minute, respectively. After air drying, the slide is finally covered by 6-diamidino-2-phenylindole (DAPI) for counterstaining nuclei with the cover slip.

After applying these four processing steps, each of the four selected Pap-smear specimen slides is marked by the two specific FISH labeled biomarkers. We asked an experienced cytogeneticist in our cytogenetic laboratory to visually and randomly select a total of 200 regions of interest (ROIs) (50 per case), which should cover the diverse image patterns typically observed in the clinical diagnostic process. Specifically, during examining each case (specimen slide) under an Olympus BX6 fluorescent microscope, the cytogeneticist manually scans the specimen slide and adjusts microscope objective focus to capture FISH signals that are considered detectable in various focus planes. Figure 1 displays a diagram of a typical microscope fluorescence imaging system, which includes a fluorescence microscope, a detector, a computer system, a motorized stage, excitation filters, emission filters, and dichroic filters. In our microscope, we installed an oil immersion-based objective lens with the magnification of 100X and a numerical aperture of 1.25 as well as a cooled charge-coupled monochrome digital camera to capture the digital images. The image spatial resolution is 0.2 μm × 0.2 μm per pixel that is adequate to capture the FISH signals, since the size of FISH spots are usually in the range of 0.5 μm~ 1 μm depending on applied probes (12, 17, 22). To avoid the registration shift, an excitation filter wheel with multi-pass dichroic and a set of barrier filters are also installed on our fluorescent microscope system. By employing separate excitation filters for each color or spectrum, the excitation filter wheel is effectively cooperating with multi-pass dichroic and barrier filters. Excitation filters allow the transmission of selected wavelengths to efficiently excite particular fluorochromes or fluorophore such as DAPI (indicating interphase nuclei), the fluorescence isothiocyanate (FITC) (marked for chromosome X), and Texas red fluorochrome (marked for chromosome 3). Thus, those fluorochromes which emit different wavelengths including spectrum blue for DAPI, spectrum green for chromosome X, and spectrum orange for chromosome 3, are passed by the cooperation of dichroic filters and barrier filters. Three digital images of each selected ROI are captured by the digital camera. Figure 2 shows an example of an acquired ROI depicting two FISH-labeled interphase cells. In all four of these selected cases, a high risk for human papillomavirus (HPV) infected lesions was detected by the cytogeneticists. Among them, two cases were diagnosed as LSIL and the other two were rated as HSIL cases; these were later diagnosed and confirmed through biopsy as having cervical cancer.

Figure 1.

Figure 1

Open in a new tab

The diagram of a FISH fluorescence microscope system.

Figure 2.

Figure 2

Open in a new tab

A FISH image of Pap-smear specimen obtained with a 100X objective lens.

An automated FISH image analysis scheme

Using these manually selected 200 ROIs as the image dataset, we developed and tested an automated FISH image analysis scheme that includes five image processing and computing steps (Figure 3): (a) an iterative multiple-threshold algorithm aiming to segment all potential interphase nuclei candidates distributed in different intensity levels; (b) a rule-based classifier to identify analyzable interphase nuclei from segmenting objects; (c) a top-hat transformation process to segment FISH spots; (d) a knowledge-based classifier to recognize splitting and stringy FISH spots as well as merge these splitting spots; and (e) computation of a summary index (detection ratio of abnormal cells) based on the number of counted FISH spots. The detailed descriptions of these steps are as follows.

Figure 3.

Figure 3

Open in a new tab

The flow chart of the proposed automated FISH imaging analysis.

Segmenting interphase nuclei using an iterative multiple-threshold algorithm

Isodata algorithm developed by Ridler and Calvard (23) is a commonly accepted iterative technique for choosing one “optimal” threshold to segment targeted objects from the image background. However, since interphase nuclei (cells) are actually located in a thin 3-D space (or distributed in different focal planes), the cells acquired in the 2-D microscopic digital images often show various intensity distributions. With a simple threshold, it is difficult to achieve optimal results for correctly segmenting all the potentially analyzable interphase nuclei. For example (Figure 4), by applying the “optimal” threshold (Figure 4b) obtained by the Isodata algorithm, the segmentation result is not very satisfactory (Figure 4c). In this example, two analyzable interphase nuclei depicted inside a clustered region with high intensity levels are not segmented. To solve this issue, we developed and tested an iterative multiple-threshold algorithm. First, a pixel value histogram of a single image frame is calculated and plotted (Figure 4d). In this histogram, three operating lines (namely, D1, D2, and D3) are selected. The first line shows the global maximum peak D1 in the histogram representing the image background, which limits the maximum sensitivity in segmenting potentially analyzable interphase nuclei candidates in the lower exposure range. The second line (D2) indicates an “optimal” threshold obtained by the Isodata algorithm to separate background and the targeted objects. The third line (D3) is empirically set as 250, which is used to segment potentially over-exposed interphase nuclei candidates.

Figure 4.

Figure 4

Open in a new tab

Illustration of segmenting interphase nuclei using two different algorithms. (a) A FISH image obtained by a 100X objective lens; (b) The optimal threshold value used for segmentation in the Isodata algorithm; (c) The segmentation result using the Isodata algorithm; (d) An iterative multi-threshold algorithm proposed in this study; (e) The segmentation result using the iterative multiple-threshold method.

The scheme then automatically generates a set of five threshold values between D2 and D3 which are Ti = D3−(i−1)×(D3− D2)/4, where i= 1,2,…,5. The sixth threshold value T6 is the global maximum representing the image background. As shown in Figure 4(d), the six selected threshold values are T = 250, 203.75, 157.5, 111.25, 65, and 8, respectively in this testing ROI. The suspicious analyzable interphase cells are segmented iteratively from the higher intensity to the lower intensity. The scheme applies the first threshold T1 to generate a binary image. All pixels whose gray value is larger than T1 are set to one and the others are assigned to zero. Since some of the FISH spots may be distributed on the edge of interphase nuclei, a morphological dilation operator with a 5×5 circle kernel is performed on each segmented object to avoid under-segmentation after obtaining potential interphase nuclei objects. The segmented objects (cells) are moved from the original image buffer to a new image storage buffer. The scheme then applies the second threshold to segment the remaining regions depicted on the same original image. Any new segmented and detected analyzable cells resulting from applying this threshold are moved and saved in the same cell storage image buffer. This process is iteratively performed for all other remaining five thresholds. As a result, all analyzable cells detected by any of these six adaptively selected thresholds are saved in a new image buffer, which will be used by the computer-aided detection scheme to detect and count FISH signals. Figure 4(e) shows the final segmentation result of applying our automated scheme to this sample image. The two analyzable cells that are unable to be segmented using a single threshold method (i.e., the Isodata algorithm) are correctly segmented by our iterative multiple-threshold method.

Identifying analyzable interphase nuclei by a rule-based classifier

After initial segmentation of suspicious interphase nuclei, a morphological opening filter with a 5×5 square kernel is applied to separate adjacent (“touching” or connected) areas and delete small isolated areas. Then, a 4-connectivity component labeling algorithm (24) is applied to detect and count the segmented cells as well as compute the following features for each labeled cell:

  1. Size of the cell (S): it is computed by counting the number of pixels (N) inside the detected or segmented area.

  2. Circularity (C): it is defined as C= (N∩ NC)/N indicating the number of pixels located both inside the labeled region or object area (N) and the equivalent circle (NC). An equivalent circle is centered at the gravity center of the labeled region and has the radius computed by RC=N/π.

  3. Compactness (CP): it is defined as C P = P2/S, where P and S are the perimeter and area of the segmented and labeled region, respectively.

  4. Length ratio (L): it is computed as L= LA/SA, where LA is the length of the longest radial length between the center and the boundary of the segmented region, while SA is the radial length along the radial line that is perpendicular to LA.

Since not all initially segmented cells represent actual analyzable cells, the scheme applies a rule-based classifier (Figure 5) to identify between the analyzable interphase nuclei and other un-analyzable “cells” (i.e., overlapped cells and other noisy debris). Applying this rule-based classifier to all segmented “cells” in the storage image buffer, the scheme classifies the initially segmented “cells” into three groups: analyzable interphase nuclei, un-analyzable objects, and indeterminate regions (i.e., length ratio is between 1.5 and 3). The indeterminate regions may contain some touching analyzable interphase nuclei. These regions will be processed again by using an iterative multiple thresholds followed by a 4-connectivitiy component labeling algorithm and a rule-based classifier mentioned above_to further segment potentially analyzable interphase nuclei. Figure 6 shows an example of applying our scheme to segment and identify analyzable interphase nuclei from a single FISH image frame. It preserves all potential interphase nuclei by deleting non-interphase objects.

Figure 5.

Figure 5

Open in a new tab

Illustration of a rule-based classifier. F1 to F5 in the figure represent five classification rules: F1 – the size of an object is between 2500 and 25000, F2 – the compactness of an object is smaller than 400; F3 – the circularity of an object is smaller than 0.8; F4 – the length ratio of an object is smaller than 3; and F5 – the length ratio of an object is smaller than 1.5.

Figure 6.

Figure 6

Open in a new tab

Identification of interphase nuclei from the segmented objects by the rule-based classifier. (a) A FISH image obtained by a 100X objective lens; (b) Initial segmentation results using iterative multi-threshold method and the computed features of each segmented region (“cell”); (c) A final segmentation result after applying the rule-based classifier.

Segmenting FISH spots by a top-hat transform

To detect and segment FISH signals depicted on the segmented interphase nuclei, our scheme separately applies the top-hat transform on two spectrums to identify and segment the corresponding FISH labeled biomarkers represented by red or green spots. The method can be described as: top− hat (f, B)= f− f◦B, where f is the original FISH image, B is a square structure with 7×7 window size, and ◦ is the morphological opening operator. Figure 7 displays the segmentation results of FISH spots by the top-hat transform. After applying the top-hat transform to the original FISH image, a threshold is applied to detect isolated FISH spots. This threshold was empirically selected based on our experimental observation and analysis of a set of randomly selected ROIs. The threshold value was selected as 100 for both red and green spectrums in this study. The scheme generates a binary image buffer for each spectrum by setting the pixel values larger than the threshold as “1” and others as “0”. Then, a 4-connectivity component labeling algorithm (24) and a raster scanning method are implemented to identify and count the number of potential FISH spots. According to the positions of FISH spots and interphase nuclei, the relationship between FISH spots and interphase cells need to be determined. The scheme detects all FISH-labeled spots that are either fully or partly located inside the detected and segmented analyzable interphase nuclei (cells) and discards all other similar “FISH” spots located on the image background.

Figure 7.

Figure 7

Open in a new tab

The segmentation results of red and green FISH spots by top-hat transformation including (a) the original FISH image; (b) the segmentation result of red FISH spot; and (c) the segmentation result of green FISH spot.

Counting independent FISH-labeled spots

Because of the characteristics of centromeric enumeration probes CEP3 and CEPX, most of FISH labeled signals used to target chromosomes 3 and X are stringy and splitting. For example, Figure 8 displays several typical types of FISH labeled signals. The red FISH signals depicted in Figure 8(a)–(b) are both split and the red FISH spots described in Figure 8(c) are stringy. Although these are all normal cells including two chromosome 3s and two chromosome Xs (represented by two red and two green FISH-labeled biomarker spots), our automated scheme detects more than two FISH-labeled spots in these cells (i.e., 3 or 4 red spots). Due to the splitting FISH spots, it may result in the wrong diagnosis of testing samples by using FISH technology. To solve this problem or minimize potential errors, an additional step is required in the automated scheme to detect and merge split FISH signal spots. Based on the hypothesis that FISH signals in the same interphase nuclei should have similar shapes and sizes, the scheme applies a pre-optimized knowledge-based classifier to identify and merge the splitting FISH signal spots (25). In brief, this classifier uses six image features that are related to the average (or relative) pixel value (intensity) inside one FISH spot, the size and shape factor of the FISH spot, and the distance between the two nearest neighbor FISH spots in the same color, to build a decision-tree type classifier. Using this classifier, the scheme enables us to (a) delete the artificial noise and/or debris, and (b) identify the splitting FISH spots. Then, the scheme merges or combines the identified splitting and stringy FISH spots and re-counts independent FISH-labeled biomarker spots depicted on each detected analyzable interphase cell.

Figure 8.

Figure 8

Open in a new tab

Examples of stringy, splitting, and typical FISH spots. (a): An illustration of different types of FISH spots; (b) – (c): Interphase cells with red splitting FISH spots; (d) – (e): Interphase cells with red stringy FISH spots.

Based on the counting results, the scheme classifies all detected interphase cells into one of two groups, the normal and abnormal (monosomy and polysomy) group. The normal cells must depict four FISH labeled biomarker spots namely two for chromosome 3 (red spots) and two for chromosome X (green spots); while the abnormal cells can be either monosomy cells (depicting only one FISH-labeled spot in any of two biomarkers) or polysomy cells (depicting more than two FISH labeled spots in any of two biomarkers). Finally, the scheme computes a detection or diagnostic summary index, which is the ratio computed based on the number of abnormal cells divided by the total number of detected cells. The detection ratios can range from 0 to 1. The larger detection ratio indicates the higher fraction of abnormal cells depicted on the testing specimen and thus represents the higher likelihood of the testing case being associated with cervical cancer.

Results

Figure 9 shows examples of applying this scheme to segment analyzable cells and detect independent FISH spots. The automated scheme segments and detects two normal and two abnormal interphase cells depicted on this FISH labeled image frame (ROI). Tables I and II summarize and compare the number of cells detected in each of the three categories (namely monosomy, normal, and polysomy) using the automated scheme and cytogeneticists’ visual detection for two FISH-labeled biomarkers targeted to chromosomes 3 and X, respectively. The automated scheme segmented and detected 73, 87, 89, and 99 analyzable interphase nuclei (cells) in four testing cases, while the cytogeneticists detected and counted 74, 92, 96, and 101 analyzable cells from the same four testing cases, respectively. The results indicated that the automated scheme missed a few analyzable FISH-labeled cells ranging from 1 to 7 cells (or 1.3% to 7.3%). As a result, the overall agreement between the automated and manual analysis ranged from 92.7% to 98.7% for the cell segmentation. Among the detected analyzable interphase cells, the automated scheme generally counted and classified more cells as abnormal cells (the number of FISH-labeled biomarker spots is unequal to two) than visual detection. The error is primarily caused by being unable to detect and count a small fraction of stringy and vague (i.e., low contrast) FISH-labeled spots. As a result, more “monosomy” cells were detected by the automated scheme as shown in Tables I and II. Such an error is larger in the LSIL cases than the HSIL cases.

Figure 9.

Figure 9

Open in a new tab

An example of automated segmentation and identification of FISH labeled spots including (a) the original FISH image; (b) the segmentation result of interphase nuclei and FISH spots; and (c) the final analysis result.

Table I.

Comparison of results in detecting FISH-labeled biomarkers for chromosome 3 (red FISH spots) in four Pap-smear examinations between the automated scheme and cytogeneticists

Case Category Automated detection Visual detection
Monosomy Normal Polysomy Monosomy Normal Polysomy
1 LSIL 11 69 3 4 74 6
2 LSIL 19 67 1 10 81 1
3 HSIL 39 39 11 44 42 10
4 HSIL 21 42 36 24 42 35
Open in a new tab

Table II.

Comparison of results in detecting FISH-labeled biomarkers for chromosome X (green FISH spots) in four Pap-smear examinations between the automated scheme and cytogeneticists

Case Category Automated detection Visual detection
1 LSIL 19 63 1 10 73 1
2 LSIL 21 64 2 11 78 3
3 HSIL 48 35 6 47 42 7
4 HSIL 19 48 32 21 46 34
Open in a new tab

Since two FISH-labeled biomarkers were used in these testing specimens, a true normal cell should contain two FISH-labeled spots in both red and green spectrums. Otherwise, the cell is classified as abnormal even if it depicts two FISH-labeled spots in one spectrum. Based on this classification criterion, Table III summarizes the total number of abnormal (including both monosomy and polysomy cells) and normal cells detected in each case as well as the final detection or diagnostic summary index. Table III compares the classification results between the automated scheme and the cytogeneticist’s visual detection. The results show that the detection ratios of two HSIL cases are substantially higher than the ratios of two LSIL cases for both automated and visual detections. Despite the differences on the classification scores, both automated and visual detection methods correctly classified these four testing cases into HSIL and LSIL categories (i.e., using the threshold of detection ratios of 50% to divide two categories). In addition, the study also found that the classification results between automated and visual detection are much more consistent or comparable in HSIL cases (i.e., difference < 4.4%, 60.7% versus 56.3%) than LSIL cases (i.e., difference < 11%, 24.1% versus 13.1%).

Table III.

Summary of final detection and classification results combining two FISH-labeled biomarkers for chromosomes 3 and X in four Pap-smear examinations between the automated scheme and cytogeneticists

Case Category Automated detection Visual detection
Normal Cells Abnormal cells Detection ratio Normal cells Abnormal cells Detection ratio
1 LSIL 63 20 24.1% 73 11 13.1%
2 LSIL 64 23 26.4% 78 14 15.2%
3 HSIL 35 54 60.7% 42 54 56.3%
4 HSIL 42 57 57.6% 42 59 58.4%
Open in a new tab

Discussion

FISH labeled biomarker image analysis has been proven to be a powerful method to detect genetic aberrations and assist the diagnosis of a variety of cancers including but not limited to cervical cancer (2), breast cancer (26), urothelial carcinoma (27), leukemia (28), and lung cancer (29). Despite of its significant potential, the FISH analysis used in current clinical practice has several disadvantages. First, it requires each observer (i.e., either a laboratory technician or a cytogeneticist) to visually identify and select approximately 100 analyzable interphase cells, which contain countable FISH-labeled biomarker spots under a fluorescent microscope. As a result, the random selection of a limited number of cells for analysis and the tendency of the observers towards the selection of cells with good morphologies will generate the inevitable error or bias. This manual selection and screening can reduce (a) diagnostic sensitivity, in particular in heterogeneous cases because the diagnostic results heavily depend on the actual searched regions and the number of cells being counted, and (b) the accuracy in quantifying residual disease in response to therapy (28). Although to the best of our knowledge, no previous studies have reported the accuracy rate of applying manual FISH technology to detect cervical cancer, the relative inaccuracy rates were reported when applying the manual FISH analysis method to detect other types of cancers. For example, one study reported an inaccuracy rate higher than 20% for breast cancer detection (26). Due to the inter-reader variability, manual FISH diagnostic accuracy also varies among different laboratories. A previous study evaluated concordance between several laboratories in the diagnosis of 813 breast cancer patients managed in the NCCTG (North Central Cancer Treatment Group) trial of adjuvant Trastuzumab therapy and found that the average concordance of FISH testing was 88% (30). Another study reported that discordance rates varied from 3% for the high-volume labs (≥ 1000 tests per month) to 24% for the low-volume labs (31).

Besides inter-reader variability, manual FISH analysis is tedious and time-consuming. For example, one study reported that it typically takes approximately 30 to 60 minutes to search for and count FISH signals of 100 cells in urine samples (32). Manual FISH analysis also cannot be feasibly applied to detect rare abnormal cells indicating early cervical cancer or other cancers in the busy screening environment (25). Therefore, to increase detection accuracy and reduce heavy workload of applying FISH technology in cancer detection and diagnosis, developing automated FISH image analysis systems or schemes has been attracting wide research interest in recent years (25). In this preliminary study, a new automated scheme was developed and tested to detect and analyze FISH-labeled biomarker spots, which target the particular changes of chromosomes 3 and X for obtaining more accurate and efficient diagnosis of cervical cancer.

To achieve higher performance, three unique approaches were tested and implemented in our automated scheme. First, to segment as many analyzable interphase cells as possible (for the maximum detection sensitivity), we proposed and implemented an iterative algorithm with multiple-stage threshold values. For those FISH images captured in the clinical environment, interphase cells are usually distributed in different intensity levels with various morphologies. Compared with the identification of a single “optimal” threshold and lower sensitivity in segmenting interphase nuclei (69%~89%) (20, 21), our proposed algorithm provides a more practical method to segment all potential interphase cells with a higher sensitivity (i.e., 92.7% in this study) and provides a foundation to achieve higher accuracy in detecting FISH-labeled biomarker spots depicted on the testing cases, which can ultimately result in higher sensitivity in detecting cancer cases.

Next, due to image noise generated during the hybridization process, a rudimentary set of image features of the initially segmented “cells” is selected and computed in this paper. Then, a set of simple knowledge-based rules is optimized and applied to identify analyzable interphase cells and delete the “non-analyzable” cells as well as other image “debris.” Instead of using a simple distance-based rule to detect and merge splitting FISH-labeled spots (20), another knowledge-based classifier is developed and applied to improve accuracy in detecting and recognizing splitting as well as stringy FISH signals depicted on the segmented cells, which is important for analyzing the true number of FISH signals and identifying abnormal cells to reduce false-positive and false-negative rates. Compared with previous research of false-positive detection rates (i.e., 7% ~ 18.8% reported by Kajtár (20) and Vrolijk (12)), our algorithm is competitive to detect abnormal cells with the maximum of an 11% false positive rate in detecting more difficult LSIL cases. As a result, this study demonstrated very comparable FISH-labeled biomarker spot counting results and overall diagnostic results using the summary index (indicating the likelihood of the testing case being associated with cervical cancer or malignant lesions) between using our automated scheme and visual examination by the cytogeneticists.

The primary limitation of this preliminary study is the small testing dataset (four testing cases) used to test the automated scheme and compare its results with manual visual detection and diagnostic results of cytogeneticists. It is well-known that in clinical practice, the accuracy of manual and visual analysis of FISH labeled images heavily depends on the quality of FISH labeled slide preparation. For example, if the slides are not washed well, the background of the FISH images becomes very noisy, which will make the clinicians unable to correctly detect the analyzable interphase cells (in particular, for those with relatively low contrast) as well as increase both false-positive and false-negative detection rates due to the increase in likelihood of miscounting or misclassifying the split and string FISH-labeled spots. We believe that the performance of any automated schemes (including ours) will also be affected by the quality of FISH slide preparation (or the noise level of FISH-labeled images). Despite this limitation, we believe that the results of this preliminary study are valid. These four testing cases were randomly selected from the FISH-labeled image slides that were actually used for clinical diagnosis and represent the average quality of FISH slide preparation in the clinical environment. The higher correlation or consistency between our automated scheme and cytogeneticists’ visual detection results is encouraging, indicating the feasibility of developing and applying the automated FISH image analysis methods. Using this data collection protocol, we are conducting a pilot prospective study to acquire a new FISH image database with several hundred cases underwent cervical cancer screening. After establishing this large and diverse database, we will test the performance and the robustness of the automated scheme to further demonstrate its potential clinical utility in our future studies.

Acknowledgments

This research is supported in part by grants from the National Institute of Health (CA136700). The authors would also like to acknowledge the support of the Charles and Jean Smith Chair Endowment fund and thank Mallory Martin for her editorial assistance.

References

  • 1.Fernlay J, Bray F, Pisani P, Parkin D. Cancer Incidence, Mortality and Prevalence Worldwide. IARC Press; Lyon: 2002. [Google Scholar]
  • 2.Zappacosta R. Cervical cancer screening: from molecular basis to diagnostic practice going through new technologies. Tech Canc Res Treat. 2008;7:161–174. doi: 10.1177/153303460800700302. [DOI] [PubMed] [Google Scholar]
  • 3.Larson NS. Invasive cervical cancer rising in young white females. J Natl Cancer Inst. 1994;86:6–7. doi: 10.1093/jnci/86.1.6. [DOI] [PubMed] [Google Scholar]
  • 4.Marzano R, Corrado G, Merola R, Sbiroli C, Guadagni F, Vizza E, Nonno FD, Carosi M, Galati M, Sperdti I, Cianciulli AM. Analysis of chromosomes 3, 7, X and the EGFR gene in uterine cervical cancer progression. Eur J Cancer. 2004;40:1624–1629. doi: 10.1016/j.ejca.2004.03.015. [DOI] [PubMed] [Google Scholar]
  • 5.Narimatsu R, Patterson BK. High-throughput cervical cancer screening using intracellular human papillomavirus E6 and E7 mRNA quantification by flow cytometry. Am J Clin Path. 2005;123:716–723. [PubMed] [Google Scholar]
  • 6.Mian C, Bancher D, Kohlberger P, Kainz C, Haitel A, Czerwenka K, Stani J, Breitenecker G, Wiener H. Fluorescence in situ hybridization in cervical smears: detection of numerical aberrations of chromosomes 7, 3, and X and relationship to HPV infection. Gynecol Oncol. 1999;75:41–46. doi: 10.1006/gyno.1999.5522. [DOI] [PubMed] [Google Scholar]
  • 7.Heselmeyer-Haddad K, Sommerfeld K, White N, Chaudhri MN, Morrison LE, Palanisamy N, Wang ZY, Auer G, Steinberg W, Ried T. Genomic amplification of the human telomerase gene (TERC) in pap smears predicts the development of cervical cancer. Am J Pathol. 2005;166:1229–1238. doi: 10.1016/S0002-9440(10)62341-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Bryan JT, Taddeo F, Skulsky D, Jansen KU, Frain BM, Qadadri B, Brown DR. Detection of specific human papilloma virus types in paraffin-embedded sections of cervical carcinoma. J Med Virol. 2006;78:117–124. doi: 10.1002/jmv.20512. [DOI] [PubMed] [Google Scholar]
  • 9.Mehes G, Speich N, Bollmann M, Bollmann R. Chromosomal aberrations accumulate in polyploidy cells of high-grade squamous interaepithelial lesions (HSIL) Pathology Oncology Research. 2004;10:142–148. doi: 10.1007/BF03033742. [DOI] [PubMed] [Google Scholar]
  • 10.Hopman A, Theelen W, Hommelberg P, Kamps M, Herrington C, Morrison L, Speel E, Smedts F, Ramaekers F. Genomic integration of oncogenic HPV and gain of the human telomerase gene TERC at 3q26 are strongly associated events in the progression of uterine cervical dysplasia to invasive cancer. J Pathol. 2006;210:412–419. doi: 10.1002/path.2070. [DOI] [PubMed] [Google Scholar]
  • 11.Netten H, Young IT, Vliet LJ, Tanke HJ, Vroljik H, Sloos WCR. FISH and chips: automation of fluorescent dot counting in interphase cell nuclei. Cytometry. 1997;28:1–10. [PubMed] [Google Scholar]
  • 12.Vrolijk H, Sloos WCR, Rijke FMvd, Mesker WE, Netten H, Young IT, Raap AK, Tanke HJ. Automation of spot counting in interphase cytogenetics using brightfield microscopy. Cytometry. 1996;24:158–166. doi: 10.1002/(SICI)1097-0320(19960601)24:2<158::AID-CYTO8>3.0.CO;2-F. [DOI] [PubMed] [Google Scholar]
  • 13.Lerner B, Clocksin WF, Dhanjal S, Hulten MA, Bishop CM. Automatic signal classification in fluorescence in situ hybridization images. Cytometry. 2001;43:87–93. doi: 10.1002/1097-0320(20010201)43:2<87::aid-cyto1022>3.0.co;2-#. [DOI] [PubMed] [Google Scholar]
  • 14.Malka R, Lerner B. Classification of fluorescence in situ hybridization images using belief networks. Pattern Recognition Letters. 2004;25:1777–1785. [Google Scholar]
  • 15.Malpica N, Solórzano COD, Vaquero JJ, Santos AS, Vallcorba I, García-Sagredo JM, Pozo FD. Applying watershed algorithms to the segmentation of clustered nuclei. Cytometry. 1997;28:289–297. doi: 10.1002/(sici)1097-0320(19970801)28:4<289::aid-cyto3>3.0.co;2-7. [DOI] [PubMed] [Google Scholar]
  • 16.Jain R, Kasturi R, Schunck BG. Machine Vision. McGraw-Hill, Inc; New York: 1995. [Google Scholar]
  • 17.Kozubek M, Kozubek S, Lukásová E, Marecková A, Bártová E, Skalníková M, Jergová A. High-resolution cytometry of FISH dots in interphase cell nuclei. Cytometry. 1999;36:279–293. doi: 10.1002/(sici)1097-0320(19990801)36:4<279::aid-cyto2>3.3.co;2-7. [DOI] [PubMed] [Google Scholar]
  • 18.Gué M, Messaoudi C, Sun JS, Boudier T. Smart 3D-fish: automation of distance analysis in nuclei of interphase cells by image processing. Cytometry. 2005;67A:18–26. doi: 10.1002/cyto.a.20170. [DOI] [PubMed] [Google Scholar]
  • 19.Lukasova E, Kozubek S, Kozubek M, Kjeronska J, Ryznar L, Horakova J, Krahulcova E, Horneck G. Localization and distance between ABL and BCR genes in interphase nuclei of bone marrow cells of control donors and patients with chronic myeloid leukemia. Hum Genet. 1997;100:525–535. doi: 10.1007/s004390050547. [DOI] [PubMed] [Google Scholar]
  • 20.Kajtár B, Méhes G, Lörch T, Deák L, Kneifné M, Alpár D, Pajor L. Automated fluorescent in situ hybridization (FISH) analysis of t(9;22)(q34;q11) in interphase nuclei. Cytometry. 2006;69A:506–514. doi: 10.1002/cyto.a.20260. [DOI] [PubMed] [Google Scholar]
  • 21.Truong K, Boenders J, Maciorowski Z, Vielh P, Dutrillaux B, Malfoy B, Bourgeois CA. Signal amplification of FISH for automated detection using image cytometry. Anal Cell Pathol. 1999;13:137–146. doi: 10.1155/1997/297216. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Netten H, Young IT, Prinsl M, Vliet LJ, Tanke H, Vrolijk H, Sloos W. Automation of fluorescent dot counting in cell nuclei. Pattern Recognition, Proceedings of Conference A: Computer Vision & Image Processing. 1994;1:84–87. [Google Scholar]
  • 23.Ridler TW, Calvard S. Picture thresholding using an iterative selection method. IEEE Trans on Syst Man CY. 1978;8:630–632. [Google Scholar]
  • 24.Mitchell TM. Machine Learning. WCB McGraw-Hill; Boston: 1997. [Google Scholar]
  • 25.Wang X, Zheng B, Li S, Zhang R, Mulvihill JJ, Chen WR, Liu H. Automated detection and analysis of fluorescent in situ hybridization (FISH) spots depicted in digital microscopic images of pap-smear specimens. J Biomed Opt. 2008;14:1–10. doi: 10.1117/1.3081545. [DOI] [PubMed] [Google Scholar]
  • 26.Bozzetti C, Personeni N, Nizzoli R, Guazzi A, Flora M, Bassano C, Negri F, Martella E, Naldi N, Franciosi V, Cascinu S. HER2/neu amplification by fluorescence in situ hybridization in cytologic samples from distant metastatic sites of breast carcinoma. Cancer. 2003;99:310–315. doi: 10.1002/cncr.11731. [DOI] [PubMed] [Google Scholar]
  • 27.Veeramachaneni R, Nordberg ML, Shi R, Herrera GA, Turbat-Herrera EA. Evaluation of fluorescence in situ hybridization as an ancillary tool to urine cytology in diagnosing urothelial carcinoma. Diagn Cytopathol. 2003;28:301–307. doi: 10.1002/dc.10291. [DOI] [PubMed] [Google Scholar]
  • 28.Dewald GW, Wyatt WA, Juneau AL, Carlson RO, Zinsmeister AR, Jalal SM, Spurbeck JL, Silver RT. Highly sensitive fluorescence in situ hybridization method to detect double BCR/ABL fusion and monitor response to therapy in chronic myeloid leukemia. Blood. 2008;91:3357–3365. [PubMed] [Google Scholar]
  • 29.Allan JM, Hardie IJ, Briggs JA, Davidson LA, Watson JP, Pearson SB, Muers MF, Wild CP. Genetic alterations in bronchial mucosa and plasma DNA from individuals at high risk of lung cancer. Int J Cancer. 2001;91:359–365. doi: 10.1002/1097-0215(200002)9999:9999<::aid-ijc1058>3.0.co;2-e. [DOI] [PubMed] [Google Scholar]
  • 30.Perez EA, Suman VJ, Davidson NE, Martino S, Kaufman PA, Lingle WL, Flynn PJ, Ingle JN, Visscher D, Jenkins RB. HER2 testing by local, central, and reference laboratories in specimens from the north central cancer treatment froup N9831 intergroup adjuvant trial. J Clin Oncol. 2006;24:3032–3038. doi: 10.1200/JCO.2005.03.4744. [DOI] [PubMed] [Google Scholar]
  • 31.Paik S, Bryant J, Tan-Chiu E, Romond E, Hiller W, Park K, Brown A, Yothers G, Anderson S, Smith R, Wickerham DL, Wolmark N. Real-world performance of HER2 testing-national surgical adjuvant breast and bowel project experience. J Natl Cancer Inst. 2002;94:852–854. doi: 10.1093/jnci/94.11.852. [DOI] [PubMed] [Google Scholar]
  • 32.Halling KC, King W, Sokolova IA. A comparison of cytology and fluorescence in situ hybridization for the detection of urothelial carcinoma. J Uro. 2000;164:1768–1775. [PubMed] [Google Scholar]

RESOURCES