Determining features in algorithms for urinary bladder cancer detection

U prvom dijelu rada bilo je potrebno izlučiti vektor značajki slika karcinoma mokraćnog mjehura pomoću dva algoritma, SIFT (eng. Scale-Invariant Feature Transform) i SURF (eng. Speeded Up Robust Features) algoritma. Rezultati su pokazali da je pomoću SIFT algoritma izlučeno više ključnih točaka što dovodi do drugog dijela rada gdje je potrebno izlučiti značajke iz originalne slike te dvije augumentirane i pronaći značajke koje se podudaraju. Također pri usporedbi vremena potrebnog da se izluče značajke, SURF algoritam je dvostruko brži od SIFT-a. U zadnjem koraku pomoću izlučenih značajki na setu podataka, trenira se umjetna neuronska mreža. Uz optimalnu arhitekturu mreže, veća točnost i manji gubici postigli su se koristeći SIFT algoritam.In the first part of the paper, a feature vector was extracted from bladder cancer images using two algorithms, the SIFT (Scale-Invariant Feature Transform) algorithm and the SURF (Speeded Up Robust Features) algorithm. More keypoints were extracted using the SIFT algorithm, in the second part of the paper it was necessary to extract the features from the original image and the two augmented ones to find features that matched. Also when comparing the time it takes to extract features, the SURF algorithm is twice as fast as SIFT. In the last step, using the extracted features on the data set, the artificial neural network is trained. With optimal network architecture, higher accuracy and lower losses were achieved using the SIFT algorithm