Using automated machine learning to detect kidney anomalies.

Abstract

483 Background: Artificial Intelligence (AI) in Urology has been used for many conditions including benign prostatic hyperplasia (BPH), urological oncology, and kidney transplant. Many computer models use algorithms that may be intricate for urologists to implement, but automated machine learning (AML) can be used to create simple models. Here we expand the use of AI and AML in image detection of kidney tumors, stones, and unremarkable kidney from computed tomography (CT) using Google Vertex AI, a machine learning platform that allows for the building, training, and deployment of models. Methods: Google Vertex AI machine learning system was trained to perform image detection. CT Kidney images were taken from publicly available data from Kaggle, an online database and machine learning platform. Images are from multiple hospitals in Dhaka, Bangladesh. 300 CT Kidney Images were uploaded on Google Vertex AI: 100 tumors, 100 stone, and 100 normal. 240 images were used to train the model, 30 for validation, and a final 30 images for assessing the accuracy of predictions after the training phase. To comprehensively evaluate our model, CT kidney images from the Cedars Sinai Medical Center were employed for further testing. All training images lacked annotations and were solely classified as normal, stone, or tumor. Results: True positivity rate for image detection during model training was 100% for tumors, stones, and normal CTs. We further tested accuracy using Cedars Sinai patient images, using 10 tumors, 10 stones, and 5 normal. The accuracy of the AI prediction was 80%, 70%, and 100%, respectively. Conclusions: Artificial Intelligence can be useful in interpreting urological imaging even in a minimally trained system. A model such as ours may allow for rapid identification and labeling of renal masses, kidney stones, and normal studies with moderate fidelity. Further training of this model may increase accuracy.