Joint two-view information for computerized detection of microcalcifications on mammograms

Abstract

We are developing new techniques to improve the accuracy of computerized microcalcification detection by using the joint two-view information on craniocaudal (CC) and mediolateral-oblique (MLO) views. After cluster candidates were detected using a single-view detection technique, candidates on CC and MLO views were paired using their radial distances from the nipple. Candidate pairs were classified with a similarity classifier that used the joint information from both views. Each cluster candidate was also characterized by its single-view features. The outputs of the similarity classifier and the single-view classifier were fused and the cluster candidate was classified as a true microcalcification cluster or a false-positive (FP) using the fused two-view information. A data set of 116 pairs of mammograms containing microcalcification clusters and 203 pairs of normal images from the University of South Florida (USF) public database was used for training the two-view detection algorithm. The trained method was tested on an independent test set of 167 pairs of mammograms, which contained 71 normal pairs and 96 pairs with microcalcification clusters collected at the University of Michigan (UM). The similarity classifier had a very low FP rate for the test set at low and medium levels of sensitivity. However, the highest mammogram-based sensitivity that could be reached by the similarity classifier was 69%. The single-view classifier had a higher FP rate compared to the similarity classifier, but it could reach a maximum mammogram-based sensitivity of 93%. The fusion method combined the scores of these two classifiers so that the number of FPs was substantially reduced at relatively low and medium sensitivities, and a relatively high maximum sensitivity was maintained. For the malignant microcalcification clusters, at a mammogram-based sensitivity of 80%, the FP rates were 0.18 and 0.35 with the two-view fusion and single-view detection methods, respectively. When the training and test sets were switched, a similar improvement was obtained, except that both the fusion and single-view detection methods had superior test performances on the USF data set than those on the UM data set. Our results indicate that correspondence of cluster candidates on two different views provides valuable additional information for distinguishing FPs from true microcalcification clusters.

Fig. 1

The block diagram of the relationship between the single-view detection, exclusive two-view detection, and two-view detection methods. The lesion prescreening, single-view classifier, and fusion blocks for the CC and MLO views are identical.

Fig. 2

The distribution of the subtlety ratings (1: most obvious, 5: most subtle) for the microcalcification clusters in the training data set. The ratings were provided with the USF database. Note that in order to be consistent with our rating scale in which a subtle lesion has a higher rating (e.g., Fig. 4), we reversed the original ratings (and their interpretation) in the USF database for this figure.

Fig. 3

The distribution of the assessment ratings for the microcalcification clusters in the training (USF) data set. The assessment follows the American College of Radiology BI-RADS lexicon and was provided with the USF database. Since all cases had biopsy-proven malignant clusters, there were very few ratings of 2 (benign finding) and 3 (probably benign finding, short-interval follow-up suggested). A majority of the ratings were 4 (suspicious abnormality, biopsy should be considered), or 5 (highly suggestive of malignancy, appropriate action should be taken).

Fig. 4

The distribution of the subtlety ratings (1: most obvious, 10: most subtle) for the malignant and benign microcalcification clusters in the test (UM) data set. The ratings were provided by MQSA radiologists at UM.

Fig. 5

The distribution of the likelihood of malignancy ratings (1: least likely to be malignant, 10: most likely to be malignant) for the malignant and benign microcalcification clusters in the test data set. The ratings were provided by MQSA radiologists at UM.

Fig. 6

The extraction of the shape features. An ellipse is fitted to the segmented microcalcification using a moment method, and the lengths of the major axis (a) and the minor axis (b) are determined. The eccentricity feature was defined as $\sqrt{a^2-b^2}/a$, and the axis ratio feature was defined as a/b. The moment ratio feature was defined as the ratio of the smaller second moment of the shape to the larger second moment.

Fig. 7

The extraction of the features related to the microcalcification contrast. The white region C at the center represents the segmented microcalcification, and the surrounding gray region S represents the background, obtained by using a dilation operator as described in the text.

Fig. 8

Geometric pairing of the clusters detected on the CC and MLO views. For a cluster C_C1 on the CC view, the nipple-to-object distance R_c is computed. On the MLO view, any object that falls within the annular region delineated by the two concentric arcs R_c + ΔR and R_c − ΔR centered on the nipple location is paired with the cluster candidate C_C1 on the CC view. In this example, two pairs are defined: C_C1-C_M1 pair and C_C1-C_M2 pair. Although a third cluster candidate exists on the MLO view, it is not paired with C_C1 because it falls outside the defined annular region. The half-width ΔR of the annular region is determined using the training data set.

Fig. 9

The distribution of the NOD differences for the true cluster centroids on the CC and MLO views. Based on the distribution of the NOD differences for the training data set, the half-width ΔR of the annular region was selected as ΔR=26 mm.

Fig. 10

Mammogram-based resubstitution FROC curves for different values of the paired-object threshold th_p. Based on these FROC curves, the paired-cluster threshold was selected as th_p =0.

Fig. 11

Mammogram-based test FROC curves for the single-view and exclusive two-view detection methods. The exclusive two-view detection method had a very low FP rate, but could only reach a maximum sensitivity of 69%. The single-view detection method had a higher maximum sensitivity, but had a higher FP rate than the exclusive two-view detection method at low sensitivity.

Fig. 12

FROC curves for the single-view and two-view fusion methods for the entire test data set. (a) Mammogram-based, (b) Case-based.

Fig. 13

FROC curves for the malignant test set. (a) Mammogram-based, (b) Case-based.

Fig. 14

FROC curves for the benign test set. (a) Mammogram-based, (b) Case-based.

Fig. 15

Single-view and two-view fusion mammogram-based test FROC curves when the training and test sets are switched. To obtain these curves, classifiers were trained on the UM data set. The trained classifiers were then applied to the USF data set.

Fig. 16

Mammogram-based FROC curves for different values of the paired-clusters threshold th_p. When the value of th_p was set to less than that obtained by training (th_p <0) the FROC curves for the test set were higher than the test FROC curve obtained with th_p =0, which was determined by training.