UCLA algorithm nearly matches experienced radiologists in prostate cancer detection

Researchers from the University of California, Los Angeles, have developed an artificial neural network capable of identifying and diagnosing prostate cancer almost as well as radiologists with a decade of experience.

The system, FocalNet, finds and predicts the aggressiveness of cancers by evaluating MRIs, much like radiologists. Unlike radiologists, it does that by using an algorithm that comprises more than a million trainable variables.

“Multiparametric MRI (mp-MRI) is considered the best noninvasive imaging modality for diagnosing prostate cancer,” first author Ruiming Cao and colleagues wrote in IEEE Transactions on Medical Imaging. “However, mp-MRI for prostate cancer diagnosis is currently limited by the qualitative or semi-qualitative interpretation criteria, leading to inter-reader variability and a suboptimal ability to assess lesion aggressiveness.

“Convolutional neural networks are a powerful method to automatically learn the discriminative features for various text tasks, including cancer detection.”

The team at UCLA trained FocalNet by feeding it scans from 417 men with prostate cancer. The algorithm was asked to analyze the images, filing away the information it learned to better assess and classify tumors in the future. In testing against experienced radiologists who analyzed the same images, FocalNet achieved an accuracy of 80.5%—just below the radiologists’ 83.9%.

Cao et al. said their research suggests an AI system like FocalNet could save time in cancer diagnosis and might serve as a tool to provide diagnostic guidance to radiologists with less real-world experience.