HRIMAC Project

HRIMAC is designed as a content-based image retrieval CAD system. This kind of CADs aims to return similar previously diagnosed cases to a given new one. The analysis of these returned cases might help the radiologist in issuing his/her diagnosis. Thus, given a mammographic image, HRIMAC searches its database in order to provide the most related cases, according to some specific criteria, which are the shape of micro-calcification and its distribution in clusters.

As it is shown in Figure , the HRIMAC's architecture can be divided into two main blocks: an on-line block and an off-line block (the grey box). In the on-line block the radiologist is studying a new mammogram of a woman. Therefore, in a fast way, the expert wants/needs to find similar analyzed cases to this one. These similar cases are the basis of the off-line block, because they had been previously characterized, and stored into a huge database of cases.

In more detail, the system is composed by:

Figure 1.4: Typical architecture of a CAD system that retrieves similar diagnosed cases compared to an unknown one. The grey square symbolizes the off-line block, meanwhile the rest of figure belongs to the on-line block.

• Easy Interface. The interaction between the computer and the radiologist has to be as simple as possible. Moreover, the interface provides a large number of tools to help, for example, to detect abnormalities, to measure areas of interest, to compare mammograms, etc.
• Mammographic Database. A CBIR CAD system contains a database of mammograms previously diagnosed, in order to be able to return the most similar cases to the radiologist. This database constitutes the knowledge of the system. As more known (and correctly diagnosed) cases are contained in the database, the accuracy of the system is expected to improve.
• Unknown Case. The unknown case is the new case that the radiologist wants to diagnose.
• Segmentation. Segmentation refers to the detection of abnormalities in the mammogram. As shown in Figure  this is the first step in both blocks. However, while the segmentation in the on-line block has to be relatively fast, the time to segment the image in the off-line block is less important.
• Characterization. The characterization of mammograms is done according to the features extracted from the abnormalities. Note that in mammography, colour is not a discriminative feature, as the images are grey-level. Thus, typical used features are related with histogram information [,196] although texture features are nowadays gaining importance [59,201].
• Database of Characteristics or HRIMAC Database. This is the database constructed from the characterization of the database of mammograms. It is required in order to reduce the computational cost of the search comparison. Furthermore, this database is indexed in order to avoid unnecessary comparisons.
• Image Retrieval. As it is shown in Figure , the image retrieval step is referred to the interaction between both blocks. When the new case has been characterized, this block has to find similar cases in the database of characteristics, and to return the corresponding mammograms.
• Relevance Feedback. The goal of this final step is to refine the retrieved result. When the system returns similar cases to the original query it may occur that not all of them are relevant. Thus, in a second step, the user indicates to the system the interesting ones. In the next retrieval, the system (hopefully) will provide more relevant images.