The analysis of medical images is gathering, in the last years, a growing interest from the scientific community working at the crossover point among physics, engineering, and medicine. The development of computer-aided detection CAD systems for the automated search for pathologies could be very useful for the improvement of physicians’ diagnosis. A typical example is the analysis of mammographic images, which are widely recognized as the only imaging modality for an early detection of breast neoplasia. Breast cancer is reported as the leading cause of woman cancer deaths in both the United States and Europe. At present, screening programs are the best known method for an early diagnosis in asymptomatic women, thus allowing a reduction of the mortality. Screening programs are based on a double visual inspection of the mammographic images, since double reading increases the diagnostic accuracy. From this point of view, the use of a CAD system could provide valuable assistance to the radiologist. In previous studies, our group proposed a CAD scheme based on ROI localization, feature extraction, and neural network classification.11 Suspect regions were detected by searching for local intensity maxima in rings whose radius was increased until the average intensity decreased to a predefined fraction of the local maximum. The ROIs thus obtained were described in terms of statistical features like average, variance, skewness, and kurtosis of the intensity distributions at different fractions of the ROI radius. This scheme relied on a simplified and rough description of the ROI, which was modeled as a round region. An improvement was achieved by implementing a new edge-based segmentation algorithm where ROIs are defined by iso-intensity contours. In this paper we retain the improved version of the segmentation step and replace the feature set with Haralik’s one. The choice for texture-based features is justified by the successful application of such features to the detection of pathologies in medical image analysis. Comparing our approach with the previously, two main aspects should be stressed. Some algorithms lack an automatic localization of the suspicious regions, rather they make use of manually selected ROIs. Other algorithms, whose scheme includes a computerized ROI hunter, lack a large and heterogeneous database to test the performance in screening-like conditions. Both these points should be taken into account in view of the development a completely automated CAD system, which should assist the radiologists in the framework of a large scale screening program. Our CAD meets both these requirements as it fits in the more general framework of the MAGIC-5 Project Medical Application on a Grid Infrastructure Connection which focuses on the development of software tools for biomedical image analysis and their use on distributed image database by means of the GRID technologies.16,17 Image collection in a screening program intrinsically creates a distributed database, as it involves many hospitals and/or screening centers in different locations. The amount of data generated by such periodic examinations would be so large that it would not be efficient to concentrate them in a single computing center. In addition, it would linearly increase with time, and a full transfer over the network from the collection centers to a central site would likely saturate the available connections. However, making the whole database available to authorized users, regardless of the data distribution, would provide several advantages. The best way to tackle these demands is to use GRID services to manage distributed databases and to allow real time remote diagnosis. This approach would provide access to the full database from any site.
Breast
