Essay: Digital Mammogram

Abstract
Digital Mammogram has emerged as a popular screening method for detection of breast cancer and other abnormalities in human breast tissues. In this paper, the different noise filtering algorithms are compared based on their ability to restructure noise exaggerated images. The intention of these algorithms is to do away the noise from a signal that might occur through the transmission of an image. The compared noise detection techniques are Median Filter (MF), Decision Based Algorithm (DBA), Modified Decision Based Algorithm (MDBA) and Progressive Switched Median Filter (PSMF). It is found that, the progressive switched median filter has better performance with low frequency salt and pepper noise compared to the other techniques.
Keywords: MF, DBA, MDBA, PSMF, Breast Cancer, Preprocessing, Filters.
1. INTRODUCTION
Breast cancer is one of the foremost cancers among female population. About 25% of all cancers diagnosed and about 20% of lethal cancers in women are breast cancers. Breast cancer is a malignant tumor that starts in the cells of the breast (American Cancer society, 2013). A malignant tumor, also recognized as a cancerous tumor, is a group of cells that have the ability to invade the surrounding tissues of the affected area and have the possibility to spread to other areas of the body. Screening is used to find breast cancer before it can start causing any symptoms. Breast cancers found during the viewing process could still be small in size and bound to the breast tissue area. The common tackles used for finding are biopsy, mammography, Positron Emission Tomography (PET), Computed Tomography (CT) scans, miraluma breast imaging, Magnetic Resonance Imaging (MRI) and bone scan. A broad treatment track for breast cancer is radiation therapy, surgery, bone directed therapy, chemotherapy, targeted therapy, hormone therapy and systemic therapy. Cancer that forms in the tissues of breast, tubes that carry milk to the nipple benign mass do not stretch to other fraction of the body but, still may need to be detached because the local tissue may be damaged. The malignant mass can destroy neighboring tissues and spread to other part of an organ or body. Mammography is the process of using low energy x-rays (usually around 30kVp) to examine the human breast and is used as a diagnostic and a viewing tool. The goal of mammography is the early detection of breast cancer, classically through detection of characteristic masses and/or microcalcifications. Microcalcifications and masses are the two important early signs of the disease.
IMAGE DETECTION TECHNIQUES
Spotting techniques are required to recover conventional mammography and develop other imaging technologies to diagnose, detect and characterize breast tumors. High-quality mammography is the useful knowledge, presently available for breast cancer screening. Efforts to get better mammogram focus on refining the technology [1]. National Cancer Institute (NCI) was backing research to reduce the low radiation dosage of mammography and enhance mammogram image quality [2]. NCI also supports research on technologies that do not use x- rays such as MRI, ultrasound and breast-specific PET scan to detect breast cancer. The following in sequence describes the latest imaging techniques that are in use or being studied.
2.1 Ultrasound
Ultrasound, also called sonography, is one of the imaging techniques in which high-frequency sound waves that cannot be heard by humans are bounced off tissues and internal organs. Ultrasound imaging of the breast is used to differentiate between fluid-filled cysts and solid tumors. It can also be used to evaluate lumps that are hard to see on a mammogram. During an ultrasound assessment, the clinician spreads a skinny covering of lubricate jelly above the area to be imaged to improve the conduction of the sound waves. A hand-held machine called a transducer directs the sound effect through the skin towards the precise tissues. As the sound effect is reflected back from the tissues within the breast, the pattern formed by the waves creates a two-dimensional image of the breast on a computer.
2.2 Digital Mammography
Digital mammography is a method for recording x-ray images in computer code instead of x-ray film, as with conservative mammography [3]. The images are displayed on a computer monitor and can be enhanced before they are written on film. Images can also be manipulated as the radiologist can expand or zip an area. Digital mammography has several advantages over conventional mammography. The images can be stored and retrieved automatically, which makes long-distance consultation with other mammography specialists easier, because the images can be familiar to the radiologist, despite subtle difference between tissues. The improved accuracy of digital mammography decreases the number of follow up actions. Digital mammography is more effective in ruling out cancer than conventional mammography.
2.3 Computer-Aided Detection
Computer-Aided Detection (CAD) involves the use of computers to bring doubtful areas on a mammogram to the radiologist’s thought. It is used after the radiologist has done the early examination of the mammogram. In 1998, the Food and Drug Administration (FDA) approved a breast imaging device that uses CAD technology whereas others are in development. An example of a breast imaging device that uses CAD technology is the Image Checker [4]. But, in 2008 systematic review on computer-aided detection in screening mammography concluded that CAD does not have a significant effect on cancer detection rate, it undesirably increase recall rate (i.e. the rate of false positives).
MRI
In MRI, a magnet connected to a computer creates pictures of areas within the body without utilizing radiation. Each MRI produces hundreds of images of the breast benign from top-to-bottom, front-to-back and side-to-side. The images are then interpreted by a radiologist. During an MRI of the breast, the patient lies on her stomach on the scanning stand. The breast hangs into gloom or hollows in the rack, which contains coils that sense the magnetic sign. The table is stimulated into a tube-like machine that contains the magnet [5]. After an original series of images has been in use, the tolerance may be given a contrast negotiator intravenously (by injection into a vein). The dissimilarity grounds are not radioactive, it is sometimes assumed to progress the visibility of a tumor. Additional images are then taken. The complete imaging assembly takes about 1 hour.
2.5 PET Scan
PET scan creates automated imaging of chemical changes that obtain a position inside tissue. If the patient agrees, an injection of an essence that consists of an amalgamation of sugar and a minute quantity of radioactive substance is administered. The radioactive sugar can aid in locating a tumor, since cancer cells attract sugar earlier than extra tissues in the body, later than in receipt of the radioactive drug. The patient remains unmoving for 45 minutes while the drug circulates all through the body. If a tumor is in attendance, the radioactive sugar will mount up in the tumor. The PET scanner is used to spot the radiation. A computer translates this in sequence into the images that are interpreted by a radiologist.
PET scan may participate a role in formative if a breast mass is cancerous. PET scans are accurate in detecting better more destructive tumors than they are in locating tumors that are smaller than 8 mm and/or less destructive. They may also perceive cancer when other imaging technique shows normal outcome. PET scans may be cared in evaluating any production periodic disease. An NCI-sponsored clinical trial is evaluating the convenience of the PET scan results in women who include breast cancer compared with the result from other imaging and diagnostic techniques. This test is also examining the efficiency of PET scans in tracking the response of a tumor to treat.
Because breast cancer cells demeanor electricity better than ordinary breast cells and tend to have lower electrical impedance, breast tumors may appear as light white acne on the computer display.
2.6 Image-Guided Breast Biopsy Techniques
Doctors execute breast biopsy, particularly of anomalous areas that cannot be felt but can be seen on a conventional mammogram or among ultrasound. One type of needle biopsy, the stereotactic-guided biopsy, involves the precise setting of the abnormal part in three dimensions using conventional mammography. A needle is then inserted into the breast and a tissue trial is obtained. Supplementary sample can be obtained by stirring the needle within the atypical area. An extra nature of needle biopsy uses a special system, known as the breast biopsy system. The handheld version of the Mammotome received FDA approval in September 1999. A large needle is inserted into the distrustful area using ultrasound or stereotactic guide [7]. The Mammotome is then put on to the vacuum tissue from the suspicious region. Additional tissue samples can be obtained by turning the spine.
3. IMAGE PREPROCESSING
Image pre-processing can notably increase the reliability of a visual inspection [8]. It is used in mammogram orientation, label and artifact removal, mammogram enhancement and mammogram segmentation. It may also involve in creating mask for pixels with highest intensity, to reduce resolutions and to segment the breast [9]. Preprocessing images normally involve removing low-frequency background noise, normalizing the intensity of the personality particle images, removing reflections and masking portions of images.
Before any image pre-processing algorithm can be applied on mammogram, the preprocessing steps are very important in order to limit the search for abnormalities without undue influence from background of the mammogram. Digital mammograms are medical images that are difficult to be interpreted, thus a preparation phase is needed in order to improve the image quality and make the segmentation results more precise. Breast border extraction and pectoral muscle suppression are also a part of preprocessing.
3.1 Noise removing tools
The types of noise observed in the mammogram are high intensity rectangular label, low intensity label, tape artifacts etc. So the objective is to remove any of these from the mammogram images if they exist. But, in this paper the focus is only on low intensity label. The label present in the mammogram image is removed by using the blob / bounding box analysis technique. For Breast Orientation depending upon the density value, it is determined whether the mammogram is of L-MLO (Left Mediolateral Oblique) view or the R-MLO (Right Mediolateral Oblique) view. If the mammogram is of the R-MLO view then the label is removed using blob / bounding box analysis technique and if the mammogram is of the L-MLO view then first it is converted into its mirror image and then the label is removed by using blob / bounding box analysis technique.
a. Median Filtering
The median filter is a nonlinear digital filtering procedure, repeatedly used to remove noise. Such noise reduction is a distinctive pre-processing step to look up the results of shortly processing [12]. Median filtering is broadly used in digital image processing below certain situation, it conserves edges even while removing noise [13, 14]. It is particularly effective in removing ‘salt and pepper’ type cluster.
This can be computed as follows:
Let I be a monochrome (1-band) image.
Let Z define a neighborhood of arbitrary shape.
At each pixel location, p = (r, c), in I.
Select n pixels in the Z-neighborhood of p.
Sort the n pixels in the neighborhood of p.
The output value at p is L (m).
The median filter works by moving through the image pixel by pixel, replacing each value with the median value of neighboring pixels [15]. The pattern of neighbors’ is called the “window”, which slides, pixel by pixel over the full image pixel, of the total image [16]. The median is considered by first arrangement, all the pixel values from the window into mathematical order [17] and then replacing the pixel being measured with the focus (median) pixel charge.
b. Modified Decision Based Algorithm
Modified Decision Based algorithm processes the corrupted images by first detecting the inclination noise [18]. The processing pixel is checked whether it is noisy or noise free [19]. If the processing pixel lies between the maximum and minimum gray level values, then it is a noise free pixel, which left unaffected.
ALGORITHM
Step 1: Read Noisy Image.
Step 2: Select 2D window of size 3×3 with centre element as processing pixel. Assume that the pixel being processed is Pij.
Step 3: If Pij is an uncorrupted pixel (that is, 0< Pij<255), then its value is left unchanged.
Step 4: If Pij = 0 or Pij = 255, then Pij is a corrupted pixel.
Step 5: If 3/4th or more pixels in selected window are noisy then increase window size to 5×5.
Step 6: If all the elements in the selected window are 0s and 255s, then replace Pij with the mean of the elements in the window. Else go to step 6.
Step 7: Eliminate 0s and 255s from the selected window and find the median value of the remaining elements. Replace Pij with the median value.
Step 8: Repeat steps 2 to 6 until all the pixels in the entire image are processed.
c. Progressive Switching Median Filter
The Progressive Switching Median filter (PSM) is a two phase algorithm. In first phase noise pixels are recognized using fixed size window (3×3). Centre pixel is treated as corrupted pixel. In second phase above mentioned knowledge of noisy pixels are identified and it is replaced by estimated median value [20]. It uses switching schema which includes two stages of noise removing techniques:
1. Preliminarily detection of noise corrupted pixels of digital images.
2. In the first stage of processing detected the filtering of noise impulses used to gather information about image properties.
3. The second stage PSM filtering procedure is used to replace such pixels with approximately correct values.
Stage-I:
Step 1: Initialize the window size (maximum value of window size is 13X13) of the filter.
Step 2: Check whether centre pixel is noisy in selected window, if YES then go to step 3. Otherwise move centre of window to next pixel and redo step 2.
Step 3: Find the value of Zmin, Zmax and Zmed in the selected window.
Step 4: Determine if Zmed is noisy by Zmin < Zmed < Zmax . If it holds, Zmed is noise free pixel and jump to step 6. Otherwise, Zmed is noisy and go to step 5.
Step 5: Increase window size and go back to Step 3.
Step 6: Replace the centre pixel with Zmed.
Step7: Reset window size and move the centre of window to next pixel.
Step 8: Repeat the steps until all pixels are processed.
At high noise density of salt and pepper noise, some of pixels are still noisy in stage-I which are further removed by passing through the entire image by stage II algorithm.
Stage-II:
Step 1: Initialize the window size of the filter by 2×2 window.
Step 2: Find out the noise free pixels present in 2×2 window.
Step 3: Find out the mean value of the noise free pixels in selected window.
Step 4: Replace the noisy pixel by the calculated mean value in step (3).
Step 5: Repeat steps from (1) – (4) to process the entire image for removal of Salt & Pepper Noise.
In PSMF, the decision is based on a predefined threshold value. The algorithm is developed based on the following two main points: Switching scheme and Progressive methods. In this, the noisy pixel can be removed either by the median value or by the mean of the previously processed neighbouring pixel values. At high noise density the median value will be 0 or 255 which is noisy. In such case, neighbouring pixel is used for replacement.
d. Decision Based Algorithm
Decision Based algorithm, image is denoised by using a window. If the processing pixel value is 0 or 255, it is processed or else it is left unchanged. At high noise concreteness the median value will be 0 or 255 which is a noisy nearby pixel is used for replacement. This repeated replacement of neighbouring pixel produces a streaking effect. The Decision Based median filter is a two phase algorithm. In the first phase, noisy pixels are recognized using fixed size window (3×3). In the second phase, prior knowledge of noisy pixels is replaced by the middle value of sorted window pixels.
ALGORITHM
Step 1: Process the first pass, K=1 starting from i1=1 and i2=1 and moving in the forward direction.
Step 2: The local window size, W and the set of uncorrupted pixels, ψ are initialized. i.e. W==3, ψ =φ .
Step 3: The set of pixel positions within a square window W×W, centred at i = (i1, i2), is defined spatially by
Ω iw={j=(j1,j2)/i1-(w-1)/2≤j1≤i1+(w-1)/2,
i2-(w-1)/2≤j2≤i2+(w-1)/2}
where W is an odd integer not less than 3 to indicate the size of the local neighbourhood under consideration. The phase begins by analyzing the pixel-wise characteristics of the corrupted image, X in the local neighbourhood, WxW.
Step 4: If M 1<Xi <M2 then Xi is declared uncorrupted and so the pixel is retained in the corresponding pixel position of the restored image, subsequently the flag image at position i, fi is reset to ‘0’ indicating a non-impulsive position.
Step 5: The purity status of Xi cannot be concluded corrupted when Xi is found not to lie strictly in between M1 and M2. The uncorrupted pixels in the neighbourhood around Xi are caught in the impulse free pixel set, ψ i.e.
ψ = { Xi / j€ Ω iw and M1 < Xi <M2 }
Step 6: If the pixel is found corrupted and ψ being non empty then X i is once again checked for its purity by analyzing correlation with the uncorrupted pixels of the window, WxW.
4. COMPARISON PARAMETERS
4.1 Peak Signal to Noise Ratio (PSNR)
PSNR is a ratio between the maximum possible power of the signal and the power of corrupting noise. It is used to measure the quality of reconstruction of lossy compression of the image. It is an estimate human sensitivity of reconstruction quality. Signals are very wide dynamic range and usually expressed in terms of a logarithmic decibel scale. The most important advantage of this measure in case of calculation, but it does not reflect perceptual quality. When PSNR value is less, it means that image is poor quality.
The mathematical representation of the PSNR is as follows:
PSNR = 20 log 10 [MAXf/√MSE] (1)
Here, MAXf is the Maximum possible pixel value of the image and MSE is the Mean Square Error. When the pixels are represented using 8 bits per sample, then MAXf is 255. More generally, when samples are represented using linear Pulse-Code Modulation (PCM) with B bits per sample, MAXf is 2B−1. For color images with three RGB values per pixel, the PSNR is the same except the MSE, where MSE is the sum of overall squared value differences divided by image size.
4.2 Root Mean Square Error
The Root Mean Square Error (RMSE) is frequently used to measure the differences between values predicted by a model or an estimator and the values really observed. These entity differences are called residuals when the calculation is performed over the data sample used for opinion and are called prediction errors when computed out-of-sample. RMSE is a good measure of accuracy, but only to evaluate forecasting errors of dissimilar models for a particular variable and not mid variables, as it is scale-dependent.
The RMSE value is computed by using the formula given below:
RMSE= √( 1/NM Σ_ij {f(i,j)-b(i,j)}2 ) (2)
The applicability of RMSE to a broad range of situations and familiarity to the general research community makes it one of the measures of choice for measuring deviation from exact location.
4.3 Noise Standard Deviation (NSD)
Noise Variance (NV) determines the contents of the speckle in the image. A lower variance gives a “cleaner” image as more tarnish is reduced, although, it doesn’t necessarily depend on the intensity [21]. The statistical measurement for NSD is given below
NSD = √NV (3)
where NV is noise variance of an image.
A noise variance allows varying from the limits of the noise control code under specific circumstances [22] usually for a limited period of time. The variance of noise is basically its “size”. When the variance of noise becomes as large as the signal being measured, it becomes extremely hard to discern the two. The mathematical representation is given below:
NV = (∑_(r,c)^n▒(I_d (r,c)-〖NMV)〗^2 ) )/(R*C) (4)
Assume that noise within an authentic image has constant statistics and noise within the forged image is statistically different from that within the authentic image. JPEG compression or local changes in illumination can affect the stationarity of the noise within the image.
The Net Magnetization Vector (NMV) in MRI is the outline of all the magnetic moments of being hydrogen nuclei. In the absence of an external magnetic field, the individual magnetic moment is randomly oriented and since they are in hostility [23] the NMV is considered to be zero. If hydrogen nuclei are placed within a strong external magnetic field, it becomes associated within the field in one of the two directions parallel to the track of the field. In MRI the main magnetic field is termed B0.
Steps
aligned in the direction of B0 (parallel)
aligned in the opposite direction of B0 (anti-parallel)
The mathematical representation of the NMV is as follows:
NMV=(∑_(r,c)^n▒I_d(r,c) )/(R*C) (5)
Where R-by-C pixels is the size of the de-speckled image Id.
The parallel and anti parallel protons cancel each other out, only the small number of low energy protons left aligned with the magnetic field create the overall net magnetization, this difference is all that counts. The magnetic moments of these protons are added together and are referred to as NMV or the symbol ‘M’.
4.4 Equivalent Numbers of Looks (ENL)
To estimate the impaired noise level in a image is to calculate the ENL over a uniform image region [24]. Due to the difficulty in identifying uniform areas in the image, it divides the image into smaller areas of 25×25 pixels, obtains the ENL for each of the smaller areas and finally takes the average of these ENL values. The mathematical representation of ENL is given below:
ENL = NMV^2/NSD^2 (6)
where NMV is mean of an image and NSD is standard deviation of an image.
The value of ENL depends on the size of the tested region, theoretically a larger region will produce a higher ENL value than over a smaller region but it also trades-off the accuracy. The significance of obtaining ENL value helps to analyze the performance of the filter for overall region as well as in smaller regions.
5. EXPERIMENTAL RESULTS
Table 1: Comparison of different filters
(a) Image – Case 001.Left_MLO (b) Image – Case 003.Right_MLO (c) Image – Case 005.Left_MLO
Median
Filter
MDBA
Filter
PSMF
Filter
DBA
Filter
Mammogram image from DDSM database (a) Image – Case 001.Left_MLO (b) Image – Case 003.Right_MLO (c) Image – Case 005.Left_MLO after applying the four different low frequency filters.
6. COMPARATIVE ANALYSIS
Table 2: Comparison of filters with parameters
Image – Case 001.Left_MLO
FILTERS MEDIAN MDBA PSMF DBA
MSE 1.06E+04 1.86E+04 1.15E+04 1.86E+04
RMS 102.7403 136.2453 107.4368 136.2374
PSNR 56.0947 53.643 55.7064 53.6435
NSD 87.2171 108.7269 90.6486 108.7277
ENL 1.46E+04 2.91E+04 1.62E+04 2.91E+04
Fig. 1 Noise detection rate of Filters
Table 3: Simulation Results
Image – Case 003.Right_MLO Image – Case 005.Left_MLO
FILTERS MEDIAN MDBA PSMF DBA MEDIAN MDBA PSMF DBA
MSE 1.55E+04 1.90E+04 1.23E+04 1.87E+04 1.18E+04 2.13E+04 1.37E+04 2.13E+04
RMS 124.6497 138.1543 111.0736 136.8152 108.9663 146.1415 117.1180 146.0505
PSNR 54.4156 53.5222 55.4172 53.6068 55.5836 53.0340 54.9570 53.0394
NSD 101.2627 109.8040 93.3839 110.0977 91.1191 114.5678 96.7578 114.5332
ENL 2.35E+04 3.02E+04 1.74E+04 2.89E+04 1.69E+04 3.47E+04 2.00E+04 3.46E+04
The output of the pre-processed image is a noise eliminated and enhanced one, which will be used for the image classification. From the experimental results it is concluded that progressive switched median filter is best in mammogram image for noise removal and gives better performance by estimating the MSE, RMS, PSNR, NSD and ENL values.
In this paper, image filtering algorithms are performed on images to remove the different types of noise which either present in the image or at the time of capturing the image. In this work, four different image filtering algorithms are compared with the five different noise removing techniques. The performances of the filters are analyzed using the parameter values. PSNR indicate high variance and MSE gives low variance to provide high quality image. When NSD is low variance, it gives a “cleaner” image. A lower RMS indicates average square difference of the pixels throughout the image between the original images. A larger value of ENL usually corresponds to a better quantitative performance.
7. CONCLUSION
Mammography images from Digital Database for Screening Mammography (DDSM) database are used in this paper, for implementation of various filtering which helps to remove noise in the image. Medical images have various limitations such as low quality, presence of noise and human error in interpretation [25]. Digital image processing can help the pathologists to a great extent. The algorithms implemented are Median Filter, Decision Based Algorithm, Modified Decision Based Algorithm and Progressive Switched Median Filter. Particularly in parameter estimation of the signal and noise distributions is the most time consuming part of the algorithm. Finally compared to other filters, Progressive Switched Median Filter is the best filter with low frequency noise. The future work of research would be to improve de-noising along with the edges and applying the method of noise variance by knowing or unknowingly.