OPTIMIZATION OF CONVOLUTIONAL NEURAL NETWORK PARAMETERS FOR URINARY STONES CLASSIFICATION BASED ON THE ATTENUATION COEFFICIENT AND DISPERSIVE X-RAY SPECTRUM

• Main Author: Aziyus Fitri, Leni
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/49061
• Institution: Institut Teknologi Bandung

Urinary stone is caused by the accumulation of solid mineral materials in the urinary system (kidney, ureter, and bladder). Four types of urinary stones such as calcium, cystine, struvite, and uric acid (UA) can be treated from a patient with different techniques. UA stone can be eliminated via oral chemolysis while struvite stone that has a higher density than UA can be broken via extra shockwave lithotripsy (ESWL). Calcium and cystine stones are removed by using percutaneous lithotripsy (PCNL) or surgery. Consequently, it is necessary to classify urinary stones based on these types. Previous researchers have shown results of urinary stone classification using both CT scan and SEM-EDX. However, there are some limitations such as subjectivity and low accuracy. Therefore, further study is required fo future improvement, in this study are proposed, i.e. optimization of scanning and reconstruction parameters of urinary stones based on micro CT scan, classification of urinary stones according to dispersive X-ray spectrum, and optimization of convolutional neural network (CNN) parameters for urinary stones classification based on attenuation coefficient or HU. The samples of this study were thirty urinary stones with diameter between 1–4 cm that had been removed from the patients through surgery. This study consisted of two main stages including data preparation and optimization of the automated classification parameters. Data preparation was done by optimizing the parameters of the micro CT scan, reconstruction, and the classification of urinary stones according to dispersive X-ray spectrum. Optimization of the automated classification parameters was conducted by selecting the algorithm, learning rate, batch size, and number of epochs. There were three novelties of this study that including scanning and reconstruction parameters obtained based on the SNR value, classification of thirty urinary stones according to dispersive X-ray spectrum analysis, and developing the automated classification based on attenuation coefficient with high accuracy. The optimized scanning parameters (source voltage of 75 kV; current of 106 ?A; exposure time of 600 ms) obtained the highest SNR. Thirty urinary stones could be classified into four types: eleven calcium oxalate stones, seven calcium phosphate stones, eight mixed stones, and four UA stones. The highest validation accuracy and the lowest RMSE values of the automated urinary stones classification informed the optimized parameter latihan. The parameters were obtained with the gradient descent with momentum (SGDM) algorithm, the number of epochs of 30, L2 regularization of 0,1, momentum of 0,9, batch size of 32, and learning rate of 10-5. The test accuracy of the method was 0,9959 with classification error of 1,2%. The method could not classify an image of the mixed stone due to the similarity of their HU values in several pixels. The conclusion of this study was that the automated classification based on convolutional neural network has classified urinary stones accurately according to attenuation coefficient and dispersive X-ray spectrum. The method was success to classify urinary stones images into three categories, i.e: calcium, mixture, and UA.