Authors:
- Ivan Lorencin, University of Rijeka, Faculty of Engineering, Vukovarska 58, 51000 Rijeka, Croatia
- Sandi Baressi Šegota, University of Rijeka, Faculty of Engineering, Vukovarska 58, 51000 Rijeka, Croatia
- Nikola Anđelić, University of Rijeka, Faculty of Engineering, Vukovarska 58, 51000 Rijeka, Croatia
- Vedran Mrzljak, University of Rijeka, Faculty of Engineering, Vukovarska 58, 51000 Rijeka, Croatia
- Klara Smolić, Clinical Hospital Centre Rijeka, Krešimirova ul. 42, 51000 Rijeka, Croatia
- Josip Španjol, Clinical Hospital Centre Rijeka, Krešimirova ul. 42, 51000 Rijeka, Croatia; University of Rijeka, Faculty of Medicine, Braće Branchetta 20/1, 51000 Rijeka, Croatia
- Zlatan Car, University of Rijeka, Faculty of Engineering, Vukovarska 58, 51000 Rijeka, Croatia
Article type:
Original Scientific Paper
Abstract:
Bladder cancer is one of the most common malignancies in men in Croatia. It is characterized by a high recurrence rate and high metastatic potential. For this reason, accurate and timely diagnosis is needed in order to treat bladder cancer as successfully as possible. Cystoscopy as a diagnostic method shows poorer accuracy of Carcinoma in situ (CIS) diagnosis, where every fourth CIS remains undiagnosed. For this reason, the artificial intelligence-based approach is proposed. The standard approach to image classification is utilization of convolutional neural networks (CNN). Literature overview shows a possibility of using pre-defined CNN models, such as VGG-16. VGG-16, in this case, needs to be customized in order to adapt it to four-class classification problem. By using a customized VGG-16 model, high classification performances are achieved. When AdaGrad and AdaMax solvers are used, AUCmicro values up to 0.98 are achieved.
Keywords:
Bladder cancer, malignancies, recurrence rate, metastatic potential