Keywords

1 Introduction

In the last decades, image processing has been an active research field and played an important role in applications such as medical aided systems, geographical information, computer vision and a lot of areas in our daily life due to developments in computer and microscopy hardware [1].

In the medical applications, there are several types of medical imaging system, such as Magnetic Resonance Image (MRI), X-Ray and Position Emission Tomography (PET). However, we have to take into account that it is quite hard to get a clear image, because, in reality, an images are often corrupted by noise due to several factors and parameters such as illumination and Data acquisition [2, 3].

Nowadays, fluorescence microscope images are a primary source of information and have became an essential tool in biology and biomedical sciences. The main problems faced with fluorescence microscopy images are blurring and non-uniformity of contrast [4]. Hence, enhancement and de-noising are the most critical stages in image processing. The primary aim of the pre-processing stage is to reduce noise and highlight the interested information simultaneously. The removal of noise has become one of the main focus and challenging topics for image processing researchers. The main target is to reduce the noise while protecting the useful information and highlight it.

From the literature, there are some several approaches for image enhancement and de-noising that are dealing with different kind of images [5, 6]. These approaches show the difference in nature of the different images to be analyzed. Molecular MRI imaging can provide an imaging solution to the detection of cells, but the technology is expensive and the works found were only from simulation results. Some works have been done on the detection of lymphocytes in optical microscope images [7, 8]. These works are interesting as the cell images present low contrast. The authors addressed the problem by applying a median filter. The overall method seems efficient for optic microscope images but may not be suited for fluorescence images as the contrast is non-uniform in the immunotherapy image sequences and the cells to be detected are not lymphocytes.

In their work for immunotherapy [9], authors used very high resolution images. The cells are clear with very high contrast for some of the samples. Other approaches illustrate the variety of cell images and the difference in nature of these original images [10, 11]. Similarly, the contrast and resolution seem very high and will not need a pre-processing step. To fulfil this target, measurement of visual quality is a critical step for numerous images for judging and evaluating images to identify cells and count them to follow the process of healing.

This paper presents several linear and non-linear filters for image enhancement and de-noising applied on fluorescence microscope images. To objectively realize a comprehensive comparison between these filters and evaluate the best method, the Peak Signal to Noise Ratio (PSNR) and Delta Signal to Noise (SNR) values between original and filtered images have been used.

The remaining of this paper is organized as follows: Sect. 2 presents the methods and techniques that have been applied on microscopy images. The results will be shown and discussed in Sect. 3. Section 4 will conclude the paper by summarizing the findings and propose future works.

2 Enhancement Methods

In this paper, various filters have been implemented with a different set of parameters on the fluorescence microscope images. The filters are divided them into two main groups. The first group is composed of linear Filters and the second group of non-linear filters. We have analyzed the quality of the different filtering methods based on PSNR and Delta-SNR values.

2.1 Linear Filter

2.1.1 Mean Filter

Moving average or “mean” filter is a linear filter. It regularises the image by averaging the pixel values contained in a specify neighbourhood. This filter is one of the most commonly used for reducing white Gaussian noises and given by [12]:

$$Y(i,j) = \frac{1}{m.n}\sum\limits_{i = 1}^{m} {\sum\limits_{j = 1}^{n} {p\left( {i,j} \right)} }$$
(1)

where; m and n are the image size and (p) is the pixel that localized at (i, j) position. When the mask of this filter is going bigger, the resulting image will be more blurred and smoothed.

2.1.2 Gaussian Filter

The Gaussian filter is a type of “low pass filter” and is used to regularize images by averaging details to reduce the noise [13]. The Gaussian filter equation is given by:

$$G(x,y) = \frac{1}{{2\pi \sigma^{2} }}.e ^{{ - \frac{{x^{2} + y^{2} }}{{2\sigma^{2} }}}}$$
(2)

where \(x,y\) are the distance from origin in the horizontal and vertical axis respectively and \(\sigma^{2}\) is the standard deviation of Gaussian distribution. This filter is more preferment to filter white Gaussian noises.

2.2 Non-linear Filter

2.2.1 Median Filter

The Median filter is a non-linear filter used to reduce speckle noise in an image. The median value is computed from pixels in a specified neighbourhood. The pixels are sorted in function of their respective values to obtain an increasing order vector. The value of the central pixel is then replaced by selecting the middle value of the ordered vector [12]. It is formulated as follows:

$$G(x, y) = Median\left\{{f\left({x - k,y - l} \right),\left({k,l} \right)\epsilon W} \right\}$$
(3)

where x and y are position indices of the sample point considered, k and l characterizing the size of neighbour window W taken into account. \(f(x - k, y - l)\) is the vector of increasing ordered pixel values contained in the window W, \(G(x, y)\) is the point value of output image at position (x, y). This filter is very efficient in removing speckle and salt and pepper noises. As well, it is efficient for removing white Gaussian noise if the noise level is not too high.

2.2.2 Wavelet Transformation

Wavelet transform (WT) is a powerful tool and widely used in image de-noising and enhancement due to its energy compaction and multi-resolution characteristics. It overcomes some of the limitations of Fourier transform with its ability to represent a function simultaneously in time and frequency domain. Discrete Wavelet Transform (DWT) critically sampled form of WT provides most compact representation. We have applied the DWT based on Haar wavelet for noise removal [14]. Indeed, Haar wavelet transform is famous due to its simplicity and speed of computation. However, Haar wavelet transform decomposes complex signals into sum of basic functions. It consists of high pass and low pass filter. The decomposition of images consists of two parts; detailed coefficient part (high frequency component) and approximation part (low frequency component).

2.2.3 CLAHE

Contrast-Limited Adaptive Histogram Equalization [15], often abbreviated CLAHE, is a computer image processing technique used to enhance and improve the contrast of the image. CLAHE functions on small regions in the image, called tiles, rather than the entire image. Each tile’s contrast is enhanced, so it differs from traditional histogram equalization that is working on the entire image. This technique proposes advanced noise removal capabilities compared to the other filters presented.

2.3 PSNR and Delta-SNR

The PSNR and the Delta-SNR are commonly used as to measure of the quality of an image after the reconstruction. The Peak Signal to Noise ratio (PSNR) is the maximum possible power of a signal and the power of corrupting noise that effects the fidelity of its representation. The PSNR equation is given as follows:

$$PSNR = 10.\,log_{10} \left( {\frac{{Max_{I}^{2} }}{MSE}} \right)$$
(4)

where, mean squared error (MSE) is given as follows:

$$MSE = \frac{1}{m\,n}\sum\nolimits_{i = 0}^{m - 1} {\sum\limits_{j = 0}^{n - 1} {\left[ {I\left( {i,j} \right) - K\left( {i,j} \right)} \right]^{2} } }$$
(5)

Here, \(Max_{I}\) is the maximum possible pixel value of the image, I and K are the original and filtered images respectively. On the other hand, The Signal to noise ratio (SNR) is used as a measure of sensitivity. It is formulated as follows:

$$SNR = 10.log_{10} \left( {\frac{{P_{signal} }}{{P_{noise} }}} \right)$$
(6)

where; P are average power for signal and noise respectively. Both signal and noise must be measured at the same.

Hence, Delta-SNR is the SNR to SNR measure between original and filtered images.

3 Results

Cancer Research Malaysia team provides us with digital immunofluorescence image database for oral cancer research (see Fig. 1).

Fig. 1
figure 1

Original fluorescence microscopy image

Full size image

The final goal of the proposed research is to extract the beneficial information for detection and identification of the different cell types. 30 fluorescence microscope images has been analysed with the different filters outlined in Sect. 2.

First, linear filters have been applied, i.e. Mean and Gaussian, with different sizes: 3 × 3, 5 × 5, 7 × 7, and 9 × 9. The PSNR and Delta-SNR between original and enhanced images are resumed in Tables 1 and 2. The Gaussian filter seems to achieve better noise filtering compared to Mean filter (see highlighted PSNR and Delta-SNR values in Tables 1 and 2). On one hand, for the Mean filter, the bigger the mask, the more blurred the result images (see Fig. 2). Hence, a mean filter size of 3 × 3 has been chosen as best results due to the preservation of the edge locations that correspond to the border of the cells.

Table 1 Average for 30 images of PSNR and Delta SNR values for mean filter
Full size table
Table 2 Average for 30 images of PSNR and Delta SNR values for Gaussian for all mask sizes
Full size table
Fig. 2
figure 2

Filtered images a using mean filter [3 × 3] and b [9 ×  9]

Full size image

On the other hand, Gaussian filter (for all mask sizes and parameters) exhibits very small Delta-SNR values. This highlight the weakness of the Gaussian filtering for enhancement of immunofluorescence images (see Fig. 3). As a conclusion for linear filters, the Delta-SNR values obtained are very small, highlighting a very slight increase in image quality, despite very high PSNR values.

Fig. 3
figure 3

Filtered image using Gaussian filter [9 × 9], sigma = 0.9

Full size image

In contrast, Wavelet de-noising gives us reasonable values of PSNR and Delta-SNR while using a one level Haar decomposition (see Table 3). For two and three levels Haar decompositions, the result images are more blurred despite increased values for the PSNR and Delta SNR (see Fig. 4).

Table 3 Average for 30 images of PSNR and Delta SNR values for wavelet transform when Haar = 1, 2, 3
Full size table
Fig. 4
figure 4

Filtered image using wavelet transform a Haar = 1 and b Haar = 2

Full size image

Non-linear filters (i.e. Median and CLAHE) give us better results. As a comparison, the difference between Median and Mean Delta-SNR values are very small, but from the PSNR values, the median filter (see Fig. 5) seems to produce slightly better enhancement (see Table 4). The results obtained with the CLAHE filter (see Fig. 6), Delta SNR and PSNR values indicate a high increase in image quality (see Table 5). Indeed, the Delta-SNR values exhibit an increase superior to 3 dB, highlighting a power increase in resulting enhanced image superior to two times. CLAHE filter have been implemented with all alpha values ranging from 0.1 to 1 with steps of 0.1, where \(\alpha\) specify the Rayleigh distribution parameter. From Table 5, the higher the \(\alpha\) value, the higher the PSNR values. “ClipLimit” is another parameter specifying the contrast enhancement limit between tiles. A lower cliplimit value seems to produce better enhancement.

Fig. 5
figure 5

Filtered image using median filter [3 × 3]

Full size image
Table 4 Average for 30 images of PSNR and Delta SNR values for median filter
Full size table
Fig. 6
figure 6

Filtered image using CLAHE for ClipLimt = 0.01 and Alpha = 0.5

Full size image
Table 5 Average for 30 images of PSNR and Delta SNR values for CLAHE with different parameters
Full size table

4 Conclusion

In this paper, several linear and non-linear filters are applied for enhancement and de-noising of immunofluorescence images. The extended comparative study based on visual results, Delta-SNR and PSNR values shows the superiority of non-linear filters in terms of de-noising and enhancement. However, the results obtained from CLAHE filtering seems to be suitable for immunofluorescence images as it provides a better background removal, edge preservation and contrast enhancement as compared to the methods discussed above. Thus, this research work concludes that, CLAHE is best pre-processing image enhancing technique to be used for other post processing steps, i.e., automatic segmentation, identification and counting of the different cell types.