Renal Calculi Detection Using Convolutional Neural Networks

Abstract

[EN] Throughout their lives, 3 in 20 men and up to 2 in 20 women develop renal calculi at some stage. Renal calculi are solid masses that originate in the kidneys but can develop along the urinary tract. They appear when solutes from urine crystalise to form calculi. Calculus formation is related to diet, urinary tract infections and medications. In most cases, renal calculi can result in excruciating pain and agony. Moreover, although it does not lead to kidney failure, recurrent renal calculi can result in a functional loss of the kidney. With the help of convolutional neural networks, image preprocessing, data augmentation and Python, TensorFlow, and Keras libraries, we have built a classification model to detect renal calculi from abdominal CT scans. Consequently, we go through an iterative process of adjusting the models and preprocessing techniques to improve their performance. Finally, we compare the best performing model against industry-standard architectures such as VGG16, VGG19, ResNet and Xception. We conclude that our model outperforms the industry-standard CNNs, but it is not ready to become a medical application.

[ES] Los cálculos renales son masas sólidas de cristales que pueden formarse en los riñones y desarrollarse a lo largo del tracto urinario. Uno de los métodos para comprobar y diagnosticar un cálculo renal es a través de un TAC abdominal.

Por ello, hemos desarrollado un modelo que utiliza redes neuronales convolucionales para clasificar imágenes de TAC abdominal. Nuestras dos clases son cálculos detectados y no detectados. Pasamos por los pasos de preprocesamiento, aumento de datos de las imágenes, diseño y construcción de los modelos. A continuación, los revisamos y los ajustamos para mejorar su rendimiento.

Por último, los comparamos con los modelos más avanzados, como VGG16 y Xception.