Discussion

We have presented and reviewed different approaches to the automatic and semi-automatic segmentation of mammographic masses. We have described several algorithms, pointing out their specific features. Specific emphasis has been placed on the different strategies and a classification of these techniques has been proposed. We have seen that few algorithms are contour-based, probably due to the fact that masses often have not a definite one. Moreover, few algorithms take shape information into account, due again to the morphology of the masses, as masses can appear in a great diversity of shapes and sizes.

Further, we have evaluated $8$ of the most frequently used strategies. These methods have been fully evaluated using ROC and FROC analysis with a common database. The annotations, which were used as the gold standard, were provided by an expert mammographic radiologist. It should be made clear that none of the investigated approaches provides the best segmentation on all forty mass containing mammograms. When taking the best segmentation per mammogram into account the $A_z$ values for circular and spiculated masses are $96.0 \pm 3.6$ and $92.8 \pm 5.7$ , respectively. The equivalent $A_z$ values for the fatty, glandular and dense mammograms are $95.9 \pm 5.9$ , $94.7 \pm 2.8$ and $91.9 \pm 5.0$ , respectively. Both these results are as expected, since it is more difficult to detect spiculated lesions and it is more difficult to detect masses on a dense (and to a lesser extend glandular) mammographic background. The picture is more complicated with respect to the size of the annotated mass, where for the same range as used in Table  the $A_z$ values are $96.4 \pm 4.3$ , $94.5 \pm 7.4$ , $92.9 \pm 2.4$ , $94.1 \pm 3.0$ , $93.6 \pm 6.2$ , and $95.7 \pm 3.6$ , respectively. For the smallest masses this is achieved by algorithm $d1$ , whilst at the large end of the scale it is algorithm $d2$ that provides the best segmentation. For the middle size range the best segmentations are ascribed to a mix of the algorithms with a majority achieved by algorithm $b2$ . It should be clear that for all these cases the mean $A_z$ value is high, which clearly shows an improvement on the individual approaches. One logical continuation of this line of reasoning is to use a combination of segmentation results to provide an improved segmentation approach.

As shown in Section  lesion shape, size and tissue type strongly influence the performance of the algorithms. Few algorithms make use of breast tissue information. Moreover, using FROC analysis we have seen that the algorithms tend to over-segment the image, obtaining a large number of false positive regions. In our opinion, the number of false positives can be reduced by incorporating pixel neighbourhood information for tissue classification to assist in the segmentation process.

Most of the model-based algorithms require the use of a classifier which implies training the system. Currently this is becoming less of a problem as more manually segmented mammograms can be found in the public domain. It should be noted that the quality of annotations is variable and the use of mammographic images from various sources makes normalization essential.

In summary, the reviewed segmentation techniques are still in need of improvement. Results demonstrated that the pattern matching approach using mutual information is an adequate solution if the mass is small. However, it seems less adequate for larger masses, and this is due to the fact that such masses appear in a great variety of sizes and shapes. On the other hand, the classifier based approach seems to be an all-round mass segmentation approach which copes equally well with the small and the large masses. This assessment is based on both region (FROC) and pixel (ROC) based classification.

None of the studied techniques provides perfect segmented results, probably showing that more than a single mammogram have to be used to detect masses. Thus, integration of segmentation results from ipsilateral, bilateral and temporal mammograms is expected to bring improvements in the final result. Work in this field of research has generated interest in the last few years.