Posted: Friday, May 5, 2023
Simon John Christoph Soerensen, MD, of Stanford University, California, and colleagues aimed to develop an artificial intelligence (AI) model that would rapidly and accurately identify cancer foci on MRI and compare its performance against that of radiologists. The results of this study, which were presented during the American Urological Association (AUA) Annual Meeting 2023 (Abstract MP09-01) and published in The Journal of Urology, suggest the model may help to avoid unnecessary biopsies of false-positive lesions.
“In this study, we trained and tested an artificial intelligence model on a large and diverse multi-institutional cohort of MRIs,” the authors explained. “In future work, we envision comparing model performance against radiologists at different skill levels and comparing radiologist plus artificial intelligence versus radiologist performance alone.”
The investigators focused on 1,504 subjects who underwent multiparametric MRI followed by MRI-ultrasound fusion biopsy. To train the AI model, T2-weighted images, diffusion-weighted imaging, and apparent diffusion coefficient maps from 1,404 cases were used. Using these parameters, as well as contrast-enhanced images, MRI lesions were initially identified and outlined by a radiologist.
Biopsies were performed to confirm cancers. Lesions considered for training and testing the model included lesions containing Gleason scores of 3 + 3 or at least 3 + 4 in at least one targeted biopsy core; other MRI lesions were removed, but systematic cores with significant cancer were included in model training. The AI model was tested on 100 independent cases and compared against findings from trained radiologists.
Across all metrics, artificial intelligence and radiologists had similar performances. Of note, all the differences were nonsignificant. Specifically, the AI model had a negative predictive value of 0.66 ± 0.26, AUC of 0.79 ± 0.22, sensitivity of 0.34 ± 0.41, and specificity of 0.99 ± 0.05. Radiologists had a negative predictive value of 0.69 ± 0.28, AUC of 0.81 ± 0.22, sensitivity of 0.37 ± 0.45, and specificity of 0.99 ± 0.05.
Disclosure: No disclosure information was provided.