Identifying Urethral Strictures Using Machine Learning: A Proof-of-Concept Evaluation of Convolutional Neural Network Model- Beyond the Abstract

Artificial intelligence (AI) has already had innovative effects in medicine and new applications are constantly being discovered. Dr. Justin Kim et al. uses a novel approach of categorizing urethral strictures using AI in his paper “Identifying urethral strictures using machine learning: a proof‑of‑concept evaluation of convolutional neural network model”. This study demonstrates the viability of using image processing techniques to evaluate urethral strictures to ultimately aid in the decision-making process for determining appropriate surgical reconstructive options.

In the study, retrograde urethrogram (RUG) images were characterized into ‘stricture’ and ‘normal’ groups to train a convolutional neural network (CNN). CNNs are a type of artificial neural networks, which are machine learning algorithms designed to mimic how our brains learn and make decisions. CNNs consist of nodes (modeled after neurons) that are interconnected to send signals of varying strength. These nodes are organized in layers: a single image input layer, several convolutional layers, and an output layer. As RUG images are processed by the CNN, the connections that resonate with the correct predictions are strengthened while those that conflict is diminished, a process called learning. Thus, the more images are used for training, the better the network will be at predicting. To prevent the CNN from being biased with the training set on hand, the researchers collected images from numerous sources and applied random transformations to reflect the variability of images seen in practice. The final CNN was able to predict the presence of urethral stricture with 88.5% accuracy on novel RUG images.

The study’s promising results show there is potential for AI to accurately identify the presence of urethral strictures in patients, even with very limited number of training data. This is likely due to the simplicity of RUG as it is a 2D flat image that can be interpreted easily. However, there is certainly more work to be done. Future studies employing larger training sets would create models with greater predictive accuracy and sophistication. This is something the author group acknowledges and future data-sharing collaborations are invited to expand the sample size and improve model performance. Ultimately, to aid with operative planning and disease prognostication, the final versions of the model should characterize details such as length, location, and etiology of stricture, and ultimately use this information to predict treatment outcomes. This study is a proof-of-concept that demonstrates the future clinical utility of AI models — especially in a setting where there is limited access to a radiologist or urologist who can interpret such images. With that said, studies such as this will help us move towards the integration of artificial intelligence into the field of medicine to improve patient care.

Written by: Kellie Kim & Jin Kyu Kim, University of Toronto, Toronto, ON

Read the Abstract