Computer‐aided diagnosis: Automatic detection of microcalcifications on digital mammograms (in Spanish)

Abstract

A computerized scheme to detect clustered microcalcifications in digital mammograms was developed. Detection of individual microcalcifications in regions of interest was also performed. Mammograms were previously classified into fatty or dense. The wavelet basis and reconstruction levels were selected. The symlets 8 were chosen for fatty tissue, and the Daubechies 4 for dense tissue. Two methods to detect individual microcalcifications were evaluated: two‐dimensional and one‐dimensional wavelet transform. The second technique yielded the best results, and was used to detect clustered microcalcifications in the complete mammogram, employing a scheme involving: (1) detection of the breast border, (2) application of one‐dimensional wavelet transform, (3) application of gray level threshold, (4) clustering procedure, (5) reduction of false positives with discriminant analysis. When detecting individual microcalcifications, we have obtained a sensitivity of 66% and an average of 7.23 false positives per image with the two‐dimensional method, and 71% of sensitivity at an average false positive rate of 6.12 with the one‐dimensional technique. For the detection of clustered microcalcifications, the sensitivity was 80% with an average of 0.94 false positives per image for fatty mammograms. The area under the AFROC curve was For dense mammograms, 73% of sensitivity at a false positive detection rate of 2.21 per image was achieved Globally, a sensitivity of 76% and an average of 1.57 false detections per image were obtained. We conclude that a wavelet‐based method to detect clustered microcalcifications would help radiologists as a “second opinion” in mammographic screening. The low false positive rate indicates that our technique would not confuse the radiologist by suggesting normal regions as suspicious.