The study involved a total of 150 cases, including 100 cases of urinary stones and 50 cases without urinary stones. These cases were reviewed by a specialist from patients who visited Sejong Chungnam National University Hospital between July 2020 and October 2022. All personal information was removed from the collected data to protect privacy, and each case was assigned a research target number. The data were then stored separately in the DICOM format on individual storage devices. Fig. 1 is sample from the dataset.

For the data preprocessing stage, the image was standardized to enhance data quality, and its size was adjusted to 224×224 to align with the input dimensions of ResNet. Subsequently, the histogram smoothing technique [11] was employed to improve the image contrast. This method involves adjusting the distribution of pixel intensity to achieve uniformity throughout the image, utilizing histograms that display the frequency of each pixel intensity level.

The procedure for histogram smoothing was as follows. Initially, the histogram function (1) was computed, which determined the number of pixels in the input image, and then normalization (2) was performed. Subsequently, the cumulative distribution function (CDF) (3) for each intensity value was calculated. The CDF (g) represented the number of pixels below the intensity g, while p(i) denoted the probability for intensity I. Finally, the CDF was used to map each pixel value of the original image to a new value. The new pixel value, dst (x, y), was calculated by multiplying the entire intensity range by the CDF. This process aided in extracting more detailed information from images that had low contrast or poor lighting. Ultimately, this significantly contributed to enhancing the accuracy of determining the presence or absence of urinary tract stones.

The data utilized in this study comprised urinary tract stones, which were expanded through various data augmentation techniques [12]. These methods have been employed to enhance the model’s performance in processing diverse urinary images in real-world settings. Data augmentation techniques were implemented using arbitrary angle rotation, horizontal/vertical movement, proportion adjustment, image reversal, and Cycle-GAN [13] data augmentation algorithms. CycleGAN is a type of generative adversarial network (GAN) that transforms images into different styles. The crux of this model lies in maintaining consistency when converting a transformed image back into its original form, effectively transferring data from one domain to another. It has applications in various fields, such as art style transformation, photo colorization, and landscape transformation. These methods enable models to be trained with a larger number of images and extract the necessary information for feature detection. Fig. 2 displays the results of data augmentation using CycleGAN.

In the design and learning process of this model, transfer learning was employed. Transfer learning is a technique commonly used in various fields, such as image classification, natural language processing, and speech recognition. In machine learning and deep learning, it involves using the weights of a pretrained model as the initial values for a new model, which can shorten learning time and improve performance. This approach is particularly beneficial when there is a scarcity of data. Consequently, this study developed a urinary tract detection model based on the ResNet50 model, which was pretrained with the ImageNet dataset. The ResNet50 model is a deep neural network with a depth of 50 layers and a structure that incorporates residual connections to overcome learning difficulties that may arise in neural networks. Compared to other transfer learning models, ResNet50 performs exceptionally well in image classification tasks. Fig. 3 illustrates the model structure proposed in this research. The output portion of the model utilized a sigmoid activation function for the binary classification problem, which determines the presence or absence of urinary tract stones. Binary cross-entropy was employed as the loss function, and the Adam optimization algorithm was used.

We enhanced the performance of the Resnet50 model through hyperparameter tuning. Key hyperparameters include the learning rate, batch size, number of epochs, weight loss, and so forth. Both grid search and random search were employed to determine the optimal values for these hyperparameters. During this process, the learning rate was initially set at 0.001 and gradually decreased as learning progressed. The batch size was established at 16, while the number of epochs was determined based on the model’s performance convergence. To avert overfitting, L2 normalization was applied to the model.

In this study, k-fold cross-validation was employed to assess the performance of the developed models. Cross-validation involves partitioning a dataset into multiple subsets, utilizing each subset as validation data while concurrently using the remaining subsets as training data. This procedure is repeated until each subset has been used as validation data once.

Ten-fold cross-validation is a technique that involves dividing the data into 10 subsets, training the model on the remaining 9 subsets, and using each subset as validation data. This approach ensures that all data is used equally for training and validation, allowing for a more accurate estimation of the model’s generalization performance. During this process, the model’s performance is measured on each fold, and the average value of these results serves as the final performance indicator. In this study, accuracy, sensitivity, specificity, and the F1-score were used as performance evaluation indicators for the model. Accuracy represents the proportion of correct predictions among all predictions, sensitivity indicates the proportion of true positives predicted, and specificity denotes the proportion of true negatives predicted in relation to the actual speech. Furthermore, the F1 score is an indicator of the harmonic mean of precision and recall, which is commonly used to evaluate model performance in unbalanced datasets. By employing these measures, this study objectively assessed the performance of the urinary tract stone diagnosis model.