Abstarct: Cystoscopy imaging is highly recommended for the early diagnosis of bladder cancer, which is the ninth most common cancer worldwide. Unfortunately, few studies have focused on image-baxsed bladder cancer diagnosis, which could be due to the lack of a standard dataset of cystoscopy images of bladder as well as variation in rotation, scale, and translation of the prepared images. In this regard, designing an intelligent system to extract discriminative features is essential for these kind of tasks. In this thesis, two main methods have been proposed baxsed on low-level and semantic features, respectively. In the first method, modified multi-laxyer perceptron baxsed on adaptive leaning rate and weight initialization baxsed on Genetic Algorithm have been used for cystoscopic image classification. The proposed method improved the convergence speed and classification accuracy. For feature extraction, local binary patterns followed by principle component analysis have been applied on the original images. The experimental results demonstrated the outperformance of the first method compared to other conventional approaches baxsed on low-level features. However, the obtained results were not proper for real world applications. In this regard, a novel approach for cystoscopic image classification baxsed on convolutional neural network has been presented. In the presented work, transfer leaning is employed for extracting features from the collected images due to the limited number of training samples. The proposed method is robust to imaging variations. In fact, previously tuned weights by ImageNet are used in the deep learning phase. To avoid the curse of dimensionality and time-space complexity, the number of features have been decreased by feature reduction methods. Finally, an ensemble classifier has been proposed for cystoscopic image classification. In the proposed method, weighted majority vote strategy is adopted to combine a set of well-known classifiers. The proposed method achieved promising results in cystoscopic image classification. Afterwards, to improve the classification performance, regional segmentation in feature maps obtained from convolution laxyers has been used to select relevant areas. High level features and the related areas have been extracted from the Particle Swarm Optimization (PSO) and Otsu method. Then, the number of features has been reduced using PCA. At the end, an ensemble classifier including Gaussian Bayes and Support Vector Machine have been utilized baxsed on majority vote. The proposed method is evaluated on 720 cystoscopy images collected in a medical center in the Netherlands. Next, the suggested method is categorized into four different classes including bloody urine, benign masses, malignant masses, and normal cases. The results of the experiments indicated that the presented work achieved a precision of 72.5%, an accuracy of 70.5%, a recall of 71%, a specificity of 89.5%, and an F1-measure of 72.5% which outperformed other competing methods.