Resumen

ROA MARTINEZ, Sandra Milena  y  LOAIZA CORREA, Humberto. EVALUATION OF TECHNIQUES FOR RELEVANCE ANALYSIS OF RADIOLOGICAL IMAGES USING FILTERS. Rev. ing. biomed. [online]. 2011, vol.5, n.9, pp.26-34. ISSN 1909-9762.

An important and fundamental stage in the image pattern recognition is the determination of the characteristics set that best describes the image. This paper describes a further step between the image characterization and its posterior classification or image retrieval similar to a given image, known as relevance analysis. It allows reducing the dimensionality of an initial set of features to a new set with fewer dimensions that preserves the hit rate of the retrieval. The analyzed images corresponded to lung nodules of radiological plaques of thorax, available through the open access library available through the Japanese society of radiological technology. To achieve these results, characteristic selection algorithms based on different filters such as FOCUS, RELIEEF-F, and BRANCH & BOUND (B&B) were analyzed. In the case of RELIEF-F it was possible to save as much as 34% of the initial characteristics set without affecting the retrieval rate compared to when the 100% of characteristics were used. Further, the implemented algorithm achieved a superior performance to that of the original algorithm included in the validated Weka software. Likewise, a strategy consisting in weights averaging was implemented that was applied to identified characteristics when the algorithms RELIEF-F, FOCUS and B&B were used simultaneously. Such weighting scheme, allowed the averaging of each characteristic according to its contribution in the minimal set of relevant features, allowing to determinate their consistency. The weighting strategy allowed a 48% reduction in the characteristics, although the retrieval hit rate slightly decreased from 77% to 76%.

Palabras clave : Relevance analysis; Features extraction; Radiological images; Dimensionality reduction.

        · resumen en Español     · texto en Español     · Español ( pdf )