ABSTRACT
Breast cancer represents the second leading cause of cancerdeaths in women today and it is the most common type ofcancer in women. This paper presents some experimentsfor tumour detection in digital mammography. We investigatethe use of different data mining techniques, neuralnetworks and association rule mining, for anomaly detectionand classification. The results show that the two approachesperformed well, obtaining a classification accuracyreaching over 70% percent for both techniques. Moreover,the experiments we conducted demonstrate the useand effectiveness of association rule mining in image categorization.
KEYWORDSclassification, medical imaging, association rule mining,neural networks, image categorization, image mining.
1. Introduction
The high incidence of breast cancer in women, especially indeveloped countries, has increased significantly in the lastyears. Though much less common, breast cancer also occursin men 1[15, 14]. The etiologies of this disease are notclear and neither are the reasons for the increased numberof cases. Currently there are no methods to prevent breastcancer, which is why early detection represents a very importantfactor in cancer treatment and allows reaching ahigh survival rate. Mammography is considered the mostreliable method in early detection of breast cancer. Due tothe high volume of mammograms to be read by physicians,the accuracy rate tends to decrease, and automatic readingof digital mammograms becomes highly desirable. It hasbeen proven that double reading of mammograms (consecutivereading by two physicians or radiologists) increasedthe accuracy, but at high costs. That is why the computeraided diagnosis systems are necessary to assist the medicalstaff to achieve high efficiency and effectiveness. The methods proposed in this paper classify the digitalmammograms in two categories: normal and abnormal.The normal ones are those characterizing a healthy patient.The abnormal ones include both benign cases, representingmammograms showing a tumour that is not formed by cancerouscells, and malign cases, those mammograms takenfrom patients with cancerous tumours. Digital mammogramsare among the most difficult medical images to beread due to their low contrast and differences in the typesof tissues. Important visual clues of breast cancer includepreliminary signs of masses and calcification clusters. Unfortunately,in the early stages of breast cancer, these signsare very subtle and varied in appearance, making diagnosisdifficult, challenging even for specialists. This is the mainreason for the development of classification systems to assistspecialists in medical institutions. Due to the significanceof an automated image categorization to help physiciansand radiologists, much research in the field of medicalimages classification has been done recently [16, 20, 9].With all this effort, there is still no widely used method toclassifying medical images. This is due to the fact that themedical domain requires high accuracy and especially therate of false negatives to be very low. In addition, anotherimportant factor that influences the success of classificationmethods is working in a team with medical specialists,which is desirable but often not achievable. The consequencesof errors in detection or classification are costly.Mammography alone cannot prove that a suspicious area ismalignant or benign. To decide that, the tissue has to be removedfor examination using breast biopsy techniques. Afalse positive detection may cause an unnecessary biopsy.Statistics show that only 20-30 percentages of breast biopsycases are proved cancerous. In a false negative detection,an actual tumour remains undetected that could lead tohigher costs or even to the cost of a human life. Here isthe trade-off that appears in developing a classification systemthat could directly affect human life. In addition, thetumours existing are of different types.


Download full report
http://citeseerx.ist.psu.edu/viewdoc/dow...1&type=pdf