Results

Literature shows us many different ways to provide results in this research field. Firstly many proposals only tries to detect microcalcifications, others use source data of such low quality in which it’s really hard to find them, basically due to the low spatial resolution or poor sampling rate (8 bits is a common value but ). Other proposals are even more radical diagnosing regions of non negligible size, ignoring the task to identify malignancy typical features. And finally, a common habit consists on the use of ad-hoc databases or selection of a low and specific number of cases in wider databases. This later point makes very difficult to compare results with other research groups, but unfortunately it’s a common fact in this field.

Detailed works use to show an approach which includes microcalcification detection, region grouping of microcalcification and malignancy analysis of the previously built ROIs. This is the strategy which we have decided to use in this work and for which we present preliminary results in this article.

ROC analysis

In mammogram based cancer detection is commonly accepted to use Receiver Operating Characteristics (ROC) curves to provide system performances, specifically measuring the area below the curve. For a given binary classifier a ROC curve shows the sensitivity versus (1 – specificity). An alternative way to build a ROC curve is to plot the fraction of true positives (TP) versus the fraction of false positives (FP).

This technique can be applied both to human experts and computerized classifiers (FROC curves is more extended for the computerized case) and allows to evaluate diagnostic results. The first ROC applications were developed during the Second World War to evaluate the effectiveness of radar operators for Japanese aircraft detection in a noisy environment.

As previously stated, this technique is based in TP and FP counts, nevertheless don’t exist a uniform criterion about what must be considered a TP or a FP in a mammogram diagnostic task. A relaxed criterion considers a generated ROI as TP if the mammogram effectively has a malignant abnormality but a restrictive criterion require effective overlap between the generated ROI and the existing one to tag the ROI as a TP.

Due to the automated generation architecture proposed, we have little control about the growing process of a region of interest, so under certain situations a ROI can grow in an excessive manner. This fact can lead to a situation in which the generated ROI will always include the real ROI, if it exist, providing a bad FP. If the mammogram doesn’t contain malignancies at all the case is even worse, because the excessive growth can lead to a single false positive in the mammogram, hiding the pernicious growing effect. To account for this effect we have decided to use a new parameter called Weighted False Positive per image (WFPi) instead of the common FPi.

This parameter accounts not only for an excessive false positive grow effect but also for the true positive one. So if a ROI diagnosed as FP and has an area bigger than a defined value (the highest area marked in the DDSM-CALC database for a malignancy) this parameter will be increased as many times as it exceed the given size plus one. Analogously an excessively big TP ROI will contribute to this parameter with its exceeding area. Mathematically we can state this parameter as follows:

We are aware of the use of this parameter instead of common true positives and true negatives deteriorate our overall statistics, but we firmly consider this is the most correct way.

Result summary

In this paragraph we present some results obtained in the developments already done, as this is a work in progress they must be considered as preliminary results. At the end of this page we provide a link to a website in which we provide an extended version of these results for different established configurations.

Firstly we show overall results obtained for one of the best configuration tested.

Below we can see for the same system a ROC curve obtained for the overlay set (global, for CALC_0 subset and for CALC_1 subset):

And finally we can see system performances for all mammograms included in the BCRP_CALC set provided by DDSM developers. In this case a FROC curve is shown.