Automatic Acquisition of Immunofluorescence Images ... - IEEE Xplore

1 downloads 0 Views 313KB Size Report
Algorithms and Evaluation. Paolo Soda, Amelia Rigon, Antonella Afeltra, Giulio Iannello. Università Campus Bio-Medico di Roma, Via Longoni 83, 00155, Roma ...
Automatic Acquisition of Immunofluorescence Images: Algorithms and Evaluation Paolo Soda, Amelia Rigon, Antonella Afeltra, Giulio Iannello Università Campus Bio-Medico di Roma, Via Longoni 83, 00155, Roma, Italy { p.soda, a.rigon, a.afeltra, g.iannello }@unicampus.it Abstract In this paper, we report our experience in the development of a system for automatic acquisition of Immuno-Fluorescence Assay (IFA) images. We focus on two basic issues. Firstly, we determine an autofocus function that can deal with photobleaching, a physical phenomenon affecting automatic acquisition of IFA images, and present a set of experiments on real images that confirm its effectiveness. Secondly, we discuss if the physicians may reliably use digital IFA images in place of direct microscope observations to carry out the diagnosis. In this respect, we present the results of a preliminary experiment where physicians perform the diagnosis on a set of images both by looking directly to them at the fluorescence microscope and by looking at digital images on the screen of a workstation.

1. Introduction Connective tissue diseases (CTD) are autoimmune disorders, commonly marked by serum antinuclear autoantibodies (ANA). The recommended method for ANA testing is ImmunoFluorescence Assay (IFA) [1], [2]. The tests are examined at the fluorescence microscope to reveal the antigen-antibody reaction. In IFA diagnosis, the physician has to report two-pieces of information: fluorescent intensity and pattern description. The readings in IFA are subjected to interobserver variability which limits the reproducibility of the method. To date, the highest level of automation in IFA tests is the preparation of slides with robotic devices performing dilution, dispensation and washing operations [3], [4]. Hence, the development of a system to support physician decision may offer a solution and this is an evident medical demand [2]. In this respect, the ability to automatically and reliably acquire IFA images seems a basic milestone. In this paper, we focus on two issues related with acquisition of IFA images. The first one is the determination and validation of an autofocus procedure to cope with photobleaching, a physical phenomenon that affects

automatic acquisition of IFA images and limits the application of existing methods. The second issue is the effectiveness of automatically acquired digital images for diagnostic purposes, i.e. if the physicians may reliably use digital IFA images in place of direct microscope observations to carry out the diagnosis. In this respect we present the encouraging results of a preliminary experiment where physicians perform the diagnosis on a set of images by looking both at the fluorescence microscope and at digital images on the screen of a workstation.

2. Materials For appropriate IFA tests, current guidelines recommend the use of tumour cell line (HEp-2) substrate [1], [2] with the 1:80 titer. The images are taken by an acquisition unit consisting of the fluorescence microscope by Leica, equipped both with a 50 W mercury vapour lamp and with a monochrome CCD camera, which has squared pixels of equal side to 6.45 m. The objective is a 40x and the medium is the air. The exposure time of slides to incident light is 0.4 s. The images have a resolution of 1024x1344 pixels and a colour-depth of 8 bits; they are stored in TIFF format. We look digital images on a 19’’ flat monitor HP L1940. Monitor settings are 1280x1024 pixels and refresh rate of 60 Hz.

3. The autofocus algorithm Focus algorithms proposed in literature [5]-[10] are based on a criterion function applied to images of the same sample, acquired at different z-axis positions. The autofocus functions give a value that indicates the degree of focusing of each image. The function maximum should correspond to the optimum focus position, and it should be sharp enough to make easy its localization. In a typical focus function three different regions can be distinguished: a near-flat region, a sloped region and a quadratic region, which lies just around the function peak [9]. The algorithm is therefore organized according three subsequent phases named coarse, fine and refine, respectively.

Proceedings of the 19th IEEE Symposium on Computer-Based Medical Systems (CBMS'06) 0-7695-2517-1/06 $20.00 © 2006 IEEE

Figure 1. Image Intensity Integral. It is the sum of all pixel values in an image. Multiple images of the same area are taken at regular z-steps (the Image Integral Intensity is normalized with respect to the first value).

Since, for practical reasons, the autofocus procedure is usually started from a position moderately far from the focus, a robust autofocus function should show an evident, although limited, slope in regions far from the focus, and a sharp peak at the focus position. Many functions have been proposed in the literature. After some preliminary tests, we selected a focus function based on image differentiation [7] (referred to as {1,-1} filter in the following), and defined by the formula:

F{1,1} =   (i[ x, y]  i[ x  1, y]) x

2

(1)

y

with i(x,y) the grey level of pixel (x,y). This filter assumes that the image roughly maintains the integral image intensity (i.e. the sum of all pixels) throughout the whole autofocus procedure. Unfortunately, fluorescent molecules are affected by the photobleaching effect that bleach them in a few seconds [11]. Since multiple images at regular steps are taken as long as the focus position is adjusted moving mechanically toward the slide, integral intensity changes when acquiring IFA images (figure 1). For this reason, the {1,-1} filter does not perform well in the acquisition of IFA images in all the three phases mentioned above. Indeed, the local maximum corresponding to the focus position is not a global maximum. Analogous behaviour has been observed also for other similar functions reported in the literature [5]-[7]. To overcome such limitations, we decided both to compensate the photobleaching effect and vary the type of functions used in the three phases.

Figure 2. Shape of {1,-1} focus functions in the fine phase. The original filter is reported by a dashed line, while the compensated one is reported by a continue line.

The {1,-1} filter is a high pass filter whose value is proportional to the energy of the image. If photobleaching is not negligible, the energy decreases during the autofocus procedure, making the filter ineffective. We then decided to compensate the fluorescence decay by normalizing the {1,-1} filter with respect to the energy of the first image acquired (referred to as {1,-1} compensated filter in the following). It is defined by the following formula:

 I FIRST IMAGE 2 F{1,1} COMPENSATED FILTER = 

 F{1,1} I 

(2)

where I is the integral image intensity. The advantage of this choice is that no direct estimation of fluorescence decay is needed. This modified filter exhibits a very good performance. Figure 2 compares the shape of the original and modified filters. It is apparent that the effects of photobleaching are well compensated by the modified filter. Figure 2 refers to the fine and refine phases of the autofocus procedure. Unfortunately, when the {1,-1} compensated filter is applied in the coarse phase, it does not succeed in clearly identifying the focus region. Indeed, when the z-axis position of the image is far from the focus position, edges and borders are not observable, which makes the localization of the focus region difficult. Based on the assumption that the grey level range of the image increases as the sample came into focus [6], we found that in the coarse phase a fairly sharp peak around the interval containing the focus position characterizes a histogram-based function. It is worth noting this function exhibits poor performance in other applications [7].

Proceedings of the 19th IEEE Symposium on Computer-Based Medical Systems (CBMS'06) 0-7695-2517-1/06 $20.00 © 2006 IEEE

Figure 3. Percentages of preference for each pair of images: images focused with the proposed algorithm (black), images reported as equivalent (dashed), images manually focused (white).

The last step in defining a complete autofocus procedure was determination of the steps to use in each of the three phases. Both to improve the overall efficiency of the procedure and to minimize the required number of images, we modified a little bit the three-phase algorithm proposed in [9]. Robustness considerations related to the nature of our acquisition system motivated us to choose 500 μm instead of 100 μm as the range where moving along the z-axis. In the coarse phase z-axis is stepped by two different steps of size 60 μm and 20 μm. In the fine and refine phases, step sizes of 10 μm and 1 μm, respectively, are chosen.

4. Performance evaluation To carry out both a quantitative and a qualitative evaluation of the algorithm performance, we manually applied the algorithm to acquire the images of 15 real IFA wells with the instruments described in section 2. The quantitative evaluation concerns the total acquisition time of each image. In all tests we did, no more than 15 steps were needed to locate the focus position. Since each image requires 0.4 s of exposure time, this means that about 6 s are needed to acquire the final image. Note that, at each step, the execution time of the autofocus algorithms can be largely overlapped to the next step, and in any case it is negligible with respect to the exposure time. The qualitative evaluation concerns the subjective quality of the images obtained applying the proposed autofocus procedure. In order to carry out this evaluation, 31 testers have been chosen and asked to look at 15 pairs of images on a computer monitor. Each pair consists of a manually focused image (man-image) and an image acquired following our autofocus algorithm (auto-image). Since IFA images are sensitive to photobleaching, we randomize the

Figure 4. Differences between focus position computed by the algorithm and position manually chosen.

order in which the digital images (i.e. man-image and autoimage) are acquired. The testers are physicians, researchers and laboratory technicians, who are used to work at microscope or at computer monitor. For every pair of images, each tester selects on a subjective basis the best-focused one or, alternatively, reports that the images are equivalent. Using Student's t-statistic (p