Impact of False Positive Segmentations on Classification

In this third experiment, we examined the impact of the FP segmentations on the performance of the classifier. As seen in Figs. 13.13 and 13.14, several signals remain in the segmentation output that are not true calcifications, individual, or clusters. These signals enter the classification stage of the algorithm and are likely to affect performance. To determine the degree of this effect, we first estimated the number of TP and FP segmented clusters. This was done be comparing the segmentation output to manual outlines of the clusters and their major calcifications generated by expert mammographers. The guidelines and conventions described elsewhere [70] were followed for these estimates. Specifically, a segmented group of calcifications was considered a TP when it contained at least three segmented true calcifications [71]. A FP cluster was one that consisted of at least three segmented objects outside the area of the true cluster within a distance of <1 cm from each other. Following the above guidelines, we determined that for a 100% TP rate, an average of 2.8 FP clusters were segmented per image with the symmlet wavelet and an average of 2.0 FP clusters were segmented with the donut filter. A reduction in either FP rate was always followed by a reduction in the TP rate to levels that were not acceptable by the classification stage that, in our case, is heavily based on morphology and distribution characteristics of the individual calcifications and their clusters.

To study the impact of the FP signals on performance but without losing TP information, we did the following experiments:

(a) The 512 x 512 pixel ROIs of the 260 clusters were automatically reduced to 200 x 200 pixels. As a result, several of the edge effects and associated false signals were eliminated concentrating the analysis on the center of the region where the signal of interest (cluster) should normally be present. Both algorithm versions were applied to the reduced-size ROIs. Results suggested that the classification of both the benign and malignant cases might be improved by up to 15% for the algorithm using the symmlet wavelet filter and up to 10% for the algorithm using the donut filter. The smaller improvement in the latter case was expected because the donut filter did not show major edge effects as the symmlet wavelet did in the original ROIs (see Figs. 13.13 and 13.14). This seemed to be an easy and fast remedy to the problem of FP signals with one downside. Namely, if the clusters were off-center in the initial ROI either due to their physical location in the breast (e.g., close to the chest wall or the skin area) or due to the initial ROI selection, then important information was lost and classification could not be done successfully.

(b) In a second experiment, all FP clusters and all single, isolated false signals that were outside the boundaries of the true cluster were manually eliminated from the 512 x 512 pixel ROIs. This manual elimination was done for a subset of 30 cases that contained small calcification clusters (3-10 calcifications per cluster). The original and FP-free ROIs were then used for feature estimation and classification. The elimination of the FP signals improved the classification of both benign and malignant cases. Significant classification improvement was observed for both benign and malignant calcification clusters and both algorithm versions. Classification errors were reduced up to 30% for the benign cases and up to 50% for the malignant cases. Further analysis of these results revealed that the presence of very small false objects in the segmentation output degraded classification performance more than large false objects such as those originating from the edge artifacts.

0 0

Post a comment