H.S.Sheshadri. A.Kandaswamy
Department of ECE,PSG College of technology,Coimbatore-641004.
hssheshadri@hotmail.com

This paper deals with certain experimental investigations on computerized method of detection of malignant and nonmalignant breast masses from mammograms. The performance of this method was tested on mammograms collected from several clinics and hospitals in and around Coimbatore.This method is capable of detecting masses as small as 4.5 mm in diameter. The technoque discussed of course helps in the early detection of breast cancer, and as a consequence increases the rate of cuereness. This new method was developed on the basis of using the macro option, which enables the execution of a series of operations according to a set of parameters. These parameters were obtained using an image processing software. The principles of trial and error has been given more weightage.Experiments have been conducted on about 80mammograms to confirm the capability of the algorithm and correctly diagnose masses independently of their sources.The parameters of the algorithm were set to give the best matching ratio (defined as true positive regions, TPR) between the algorithm diagnosis and the radiologist’s diagnosis in terms of mass size and location. Percent matching was independent of mass size.This gave the algorithm high performance reliability because it indicated that its performance was insensitive to mass size. When the opinion of a second radiologist was considered, TPR was also invariable.

Keywords: Computer Aided Diagnosis,Malignant tumors, Tumor detection

Method Adopted.

The new algorithm detects small breast masses, (malignant or nonmalignant) which helps in the early detection of breast masses. A relatively large number of mammograms (table 1) were used in the development process of the algorithm. These mammograms were collected from several hospitals and clinics. This of course enriched the study and covered many conditions of breast masses.
Computerized detection of masses from mammograms can be used as a tool to reduce the number of false diagnosis, which in turn reduces number of cases transferred to biopsy. The new method is based on the following three stages;
 Knowledge collection of the nature of the mammograms.
 Obtaining radiologist’s diagnosis.
 Building the algorithm.

Table1: Sources and numbers of mammograms used in the study.

Each mammogram was labeled with three tag points, and then it was superimposed on a transparency in order for the radiologist to perform his eye-aided diagnosis. This transparency also includes the same tag points for referencing purpose as shown in figure 1.
Each mammogram and its corresponding transparency were digitized using a scanner with a 360dpi, and 256 gray level, which resulted in digital images of sizes of about 128-256KB. During the digitization process,we considered only the affected area as region of interest and stored in computer memory. ( in order to reduce the memory volume consumed by these images).
The new mass detection algorithm was built based on the following three image processing steps:
1. Brightness and contrast enhancement.
2. Histogram equalization.
3. Thresholding.
The white color was assigned to the tumor cells, and the black color was assigned to normal breast tissues and the surrounding air. Figure 2 shows a block diagram of the algorithm used. Also it gives a comparitive study of the method employed by a radiologist.

Figure1: An example of a mammogram (a), and the eye-aided diagnosis (b).

Figure 2: (A) block diagram of the radiologist’s diagnosis, (B) block diagram of algorithm.

The brightness for every image ( total of 82 images ) was increased in steps, each step equals 10 (full scale = 200). Every time the brightness was adjusted, a print out of the image was presented to the radiologist for diagnosis. The diagnosis obtained at each brightness level was compared to the original radiologist diagnosis performed on the transparency. Figure 3 shows the distribution of number of images that agrees with the original transparency-based radiologist diagnosis versus brightness level. With the distribution being close to normal, we used the distribution mean of 40 as the value to be used in the algorithm. In similar manner we obtained a value of 40 for the contrast enhancement, and 225 for the thresholding as shown in figure 4,and 5 respectively.The purpose of using histogram equalization is to normalize the distribution of the gray levels of all images. This was necessary as a result of the variety of sources and methods of mammography. In addition, it also increases the sensitivity of subsequent processing functions.

Figure 3: Distribution of images with coincided diagnosis (radiologist and algorithm diagnosis) vs. brightness level.

Figure 4: Distribution of images with coincided diagnosis (radiologist and algorithm diagnosis) vs. contrast level.

Figure 5: Distribution of images with coincided diagnosis (radiologist and algorithm diagnosis) vs. threshold level.

Results and Data Analysis

Results were presented in two forms. The qualitative results are the digitized images obtained after applying the algorithm. The quantitative results are the data analysis, and results of the statistical tests, which compares the outcome of the algorithm with the diagnosis of two radiologists.

Qualitative results
To evaluate the algorithm performance, its outcome was compared to the radiologist diagnosis using the technique of image subtraction. By defining the image outcome of the algorithm as pc (i,j), and the image outcome of the radiologist diagnosis as pd (i,j), then the subtracted image would be:Ps (i,j) = pd (i,j) - pc (i,j)
The subtracted image gives primary evaluation to the algorithm in terms of size and location of the diagnosis mass. Figure 6 shows an example of a mammogram that contains malignant mass as proven by biopsy.

Figure 6: An example of the qualitative results, (a) the original image, (b) the algorithm output, (c) radiologist diagnosis, (d) subtraction of images “b” and “c”.

Quantitative analysis
1. This analysis is based on mass location and size. These two parameters will be compared between the algorithm output image, and the radiologist diagnosis image. The concept of true positive regions (TPR), and false positive regions (FPR) will be used as a measure of the comparison between the two diagnosis. These parameters are based on the quantity of matched regions PM, total region PT, and mismatch region PMM, as described in Figure 7.

Figure 7: Concept of quantitative analysis. (a) PM, (b) PT, (c) PMM.

Where “W” is the mass size as diagnosed by the radiologist, which represents number of white pixels in the image Pd . This parameter was obtained from the image histogram.TPR and FPR will take values of 100 and 0 receptively in the ideal case, which represents total match between the algorithm diagnosis, and the first radiologist diagnosis.The algorithm parameters were based on the outcome values of TPR1 and FPR1 between the algorithm diagnosis and the first radiologist diagnosis. In this case a total of 86 mammograms were presented to the first radiologist prior to being diagnosed by the algorithm.
In order to test the independence of the algorithm performance, it was applied to a total of 23 mammograms prior to being presented to a second radiologist, in which case we obtained values of TPR2 and FPR2. Values of TPR1, FPR1, TPR2, and FPR2 are shown in tables 2, and 3

Table 2: values of TPR1, TPR2, FPR1, FPR2.

Table 3: values of TPR, FPR when masses were categorized as malignant and nonmalignant.

Discussion

The mammograms used to build this algorithm were obtained from three hospitals and clinics clinics. The algorithm was able to detect masses as small as 4.8 mm in size, which correspond to a mass age of 1-2 years according to the type of the tumor. This emphasizes the fact that this algorithm may help in the early detection of breast cancer, and as a result, increase the chance of cureness. To test the accuracy of the algorithm, 23 mammograms not used in building the algorithm were diagnosed using the algorithm prior to being presented to the first radiologist. The structure of this algorithm allows the user (i. e. the radiologist) to use the whole process, or part of it as required. For example, expert radiologist may only need to enhance brightness or contrast. The radiologists in general include a safety factor when they diagnose mammograms. This factor presents an extra area around the mass. The first radiologist suggested that this factor is about 20% of mass size. Based on that the parameters of the algorithm were chosen to give an outcome of 20% less than the radiologist diagnosis. In this case, when using the algorithm to detect masses, the radiologist may add this extra factor depending on the case All 82 mammograms were presented to the second radiologist to compare his diagnosis with that of the first radiologist. The outcome of this comparison was TPRd, FPRd. The mean , and the standard deviation , were calculated for all values of TPR, FPR for all combinations as shown in table 4.

Table 4: values of  and  for TPR, FPR.

Statistically, there was no significant differences between values of TPR1, and TPR2.

Figure 8: Mass size according to the algorithm vs. mass size according to the first radiologist.

To test the independence of accuracy of mass size, we examined the relationship between mass size according to the algorithm as a function of mass size according to the first, and second radiologist, as shown in figures 8, 9 respectively.

Figure 9: Mass size according to the algorithm vs. mass size according to the second radiologist.

The information in these figures suggests that the relationships are linear. The figures also include the linear regression lines along with the regression parameters. With values of correlation coefficient (r) larger than 0.9, it indicate a strong linear relationships. The slope of the linear regression represents the constant fractional difference between the variables. This means that the fractional difference between mass size according to the algorithm, and that according to the radiologist is independent of mass size. Figure 10 shows the relationship between mass size according to the first radiologist, as a function of mass size according to the algorithm for 23 mammograms when they were diagnose by the algorithm prior to being presented to the radiologist.
The figure also includes the linear regression lines along with the regression parameters

Figure 10: Mass size according to the first radiologist vs. mass size according to the algorithm

With the values of correlation coefficient (r) is larger than 0.9, it indicates a strong linear relationships. This indicates the capability of the algorithm to correctly diagnose mammograms that were not involved in the building process of the algorithm.
The relationship between TPR and mass size as determined by the first radiologist is as shown in
figure 11. Thus, the values of TPR are independent of mass size.

Figure 11: The relationship between TPR and mass size as determined by the first radiologist.

Conclusion

The experimental method suggested above is simple and employ direct image processing techniques to detect any abnormality in the breast tissue.The radiologists who assessed this work felt that this technoque would give better clue for early detection of breast cancer. This method can be further employed for the analysis of mammograms under a Computer Aided Diagnosis (CAD ) system.
Our future work include the development of better image processing techniques to identify microcalcifications in the breast tissue.Also the development of a CAD system for early detection of breast cancer is included in our future research work..

References
[1]ASTLEY S M and TAYLOR C J 1990 Combining cues for mammographic abnormalities Proc. 1st British Machine Vision ConferenceOxford UK 253-258
[2]CHAN H P, DOI K et al. 1987 Image feature analysis and computer-aided diagnosis in digital radiography. Automated detection of microcalcifications in mammography Medical Physics 14 (4) 538-548
[3]CHAN H P, DOI K, VYBORNY C J, LAM K J and SCHMIDT R A 1990 Improvements in radiologists' detection of clustered microcalcifications on mammograms: the potential of computer aided diagnosis Invest. Radiol. 25 1102-1110
[4]DAVIES D H and DANCE D R 1990 Automatic computer detection of clustered calcifications in digital mammograms Phys. Med. Biol. 35 1111-1118
[5]DHAWAN A P, BUELLONI G and GORDON R 1986 Enhancement of mammographic features by optimal adaptive neighbourhood image processing IEEE Trans. Med. Imag. MI-5 8-15
[6]EGAN R L Technologists guide to mammography 2nd edition Williams and Wilkins Company 1977
[7]ERIKSEN J P, PIZER S M and AUSTIN J D 1990 MAHEM: a multiprocessor engine for fast contrast limited adaptive histogram equalisation SPIE Conference Medical Imaging IV - Image Processing SPIE Vol. 1233
[8]GIGER M L, YIN F, DOI K, METZ C E, SCHMIDT R A and VYBORNY C J 1990 Investigation of methods for the computerised detection and analysis of mammographic masses SPIE Conference Medical Imaging IV - Image Processing SPIE Vol. 1233 183-184
[9]GORDON R and RANGAYAN R M 1984 Feature enhancement of film mammograms using fixed and adaptive neighbourhoods Applied Optics 23 560-564
[10]HARALICK R M, SHANMUGAN K and DINSTEIN I 1973 Textural features for image classification IEEE Trans. Sys. Man. Cyb. SMC-3 6 610-620