Wavelet transforms for detecting microcalcifications in mammography.

Abstract

A three-stage method based on wavelet transforms for detecting and segmenting calcifications is developed. The first stage consists of preprocessing which is realized by subtracting from the original image a Gaussian low-pass filtered version, followed by a full resolution wavelet transform. With a Laplacian of Gaussian wavelet basis the detection of microcalcifications can be nearly optimized in the details sub-bands. In fact, the two-dimensional filters which transform the input image into HH and LH + HL details sub-bands are closely related to prewhitening matched filters for detecting Gaussian objects in separable and non-separable first-order Markov noise, respectively. Two methods have been proposed to detect candidate microcalcifications. In the first method, the outputs HH and LH + HL from each octave are thresholded at some fixed percentile of the histogram of each component. Then, the detected images from all octaves are logically ORed to yield the binary map of detected pixels. The second method employs a Hotelling observer. The Hotelling discriminant is computed and then thresholded to obtain the binary map. The second stage is needed to reduce false alarms, which are reduced by analyzing the shapes of the detected pixel regions. The third stage is designed to provide an accurate segmentation of calcification boundaries. Individual microcalcifications are often greatly enhanced in the output image. FROC curves are computed from tests using a well-known database of digitized mammograms. The algorithm using the Hotelling observer shows the best overall performance in which a true positive fraction of 73% is achieved at 0.7 false positives per image.