Performance of Microcalcification Detection Algorithms

Abstract

Breast cancer is the most common malignancy in women and is three times more common than all gynecologic malignancies put together. The incidence of breast cancer has been increasing steadily from an incidence of 1:20 in 1960 to 1:8 women today. Seventy percent of all breast cancers are found through breast self-exams. However not all lumps are detectable by touch and mammography is a low-dose X-ray examination that can detect breast cancer up to two years before it is large enough to be felt. Some patterns of microcalcifications (small white deposits of calcium) give an early indication of cancer. Their small size makes their detection difficult for the radiologist. This brings in the role of CAD (Computer Aided Diagnosis) which serve as an assistant to the radiologist. In this thesis we have investigated the performance of three state of the art CAD techniques for the detection of microcalcifications in mammograms.

First, is a wavelet based technique which applies an adaptive wavelet filter to the input mammogram. Then it calculates HOS (Higher Order Statistics) values for maxima locations that are determined from the input image by an empirical method. This is followed by determination of a threshold using a cross entropy based thresholding algorithm. The thresholded image gives the locations of microcalcifications.

Second, is a technique that pre-processes the input mammogram with a tophat morphological filter which only preserves objects that are smaller than the size of the filter used in pre-processing. This is again followed by the determination of a threshold using the same thresholding algorithmas in the first technique. The thresholded image indicates the positions of microcalcifications. We have also done an investigation of co-occurrence matrix based entropy thresholding schemes. We found that two dimensional matrix based algorithms perform better than three dimensional based algorithms. Although both fail in case of images with high dynamic range which make them unsuitable for medical imaging. However the cross entropy based method performed better than co-occurrence matrix based techniques for both low as well as high resolution images.

Third, is a technique that makes use of a high pass filter for pre-processing. Classification of a location as a microcalcification is done by a SVM (Support Vector Machine) classifier using a scheme called SEL(Successive Enhancement Learning). These techniques have been compared by the use of ROC curves and we found out that the SVM based technique gives the lowest false positives for high detection rate. However cross entropy method does gives lower false positives for detection rates lower than 65%.