KRATKA VSEBINA | Učinkovitost segmentacijskih algoritmov za medicinske slike naj bi merili na množici kliničnih podatkov, kar pa je lahko zelo zahtevna naloga, predvsem zaradi pomanjkanja etalonov, standardiziranih statističnih protokolov, ustreznih metrik in dolgotrajnegazbiranja podatkov. V članku smo predstavili novo proceduro za lažje evalviranje učinkovitosti segmentacijskih algoritmov. Najprej sliko segmentiramo, zatem pa dobljene regije označimo. Parametrično popisane regije zatem primerjamo z originalnimi objekti. Odločitev, ali je regija razpoznana, sprejmemo na osnovi preseka med segmentirano regijo in originalnim objektom. Učinkovitost algoritmov merimo z dvema statistikama: z razmerjem med pravilno razpoznanimi objekti in regijami. Učinkovitost za celotno množico slik dobimo s poprečenjem obeh razmerij. Končno oceno o segmentacijski učinkovitosti določimo s produktom obeh vpeljanih statistik.Na osnovi predlagane evalvacijske procedure smo merili učinkovitost 10 dobro poznanih segmentacijskih algoritmov, in sicer najprej na 50 umetnih slikah, kjer se je izkazal algoritem OPTIMAL kot najboljši, ter zatem še na30 realnih posnetkih, kjer je bil najboljši algoritem UNION-FIND. // A segmentation efficiency of the algorithms for medical images should be measured on a set of clinical data, what is very demanding task, due to deficiency of golden standards, standardised statistical protocols, appropriate metrics and tedious data gathering. In this paper, a new procedure for easier evaluation of the segmentation algorithm efficiency isproposed. First, an image is segmented, afterwards, the regions obtined arelabelled. The regions described parametrically are then compared to the original objects. A decision about the region recognition is taken considering an intersection between the segmented region and the original objects. The algorithm's efficiency is measured with two statistics: the ratios of correctly recognised objects and regions. The efficiency on an entire image set is determined by averaging the ratios. A final assesment of the segmentation efficiency is determined on a product of both statistics introduced. The efficiency of ten well-known segmentation algorithms is measured according to the proposed evaluation procedure, first, on 50 artificial images where the OPTIMAL algorithm proved to be thebest, and afterwards, on 30 real images-the best was UNION-FIND. |