An artificial intelligence model combining MRI and ultrasound images (MRUSuNet) for prostate cancer detection outperforms radiologist MRI interpretation.

OBJECTIVES: To develop the magnetic resonance imaging (MRI)-UltraSound UNet (MRUSuNet), a model that uses both MRI and transrectal ultrasonography (TRUS) as input to detect and localise the three-dimensional spatial volume of clinically significant prostate cancer on TRUS images to allow better visualisation of cancer in the TRUS space. PATIENTS AND METHODS: Pre-biopsy MRI is increasingly used to target suspicious prostate lesions. The lesions identified by radiologists are mapped onto TRUS images to enable real-time targeting during the biopsy. Yet, the misalignment of MRI and TRUS results in sampling errors, and in some cases even missing the clinically significant cancers. This study included two MRI-TRUS fusion biopsy cohorts (3000 patients) for model development, and one radical prostatectomy cohort (110 patients) used for independent validation and for the comparison with radiologist interpretation. A total of 1700 independent cases were reserved to compare the MRUSuNet with single-modality MRI-only and TRUS-only models. Sensitivity and lesion-level Dice similarity scores were computed using targeted biopsy pathology as the reference standard in the biopsy cohorts and whole-mount prostatectomy pathology in the radical prostatectomy cohort. RESULT: The MRUSuNet achieved significantly superior sensitivity (0.80) and lesion Dice score (0.42) across all test cohorts compared to the MRI-only model (sensitivity: 0.73, P = 0.033; Dice score: 0.30, P < 0.001) and the TRUS-only model (sensitivity: 0.49, P < 0.001; Dice score: 0.27, P < 0.001). Compared to radiologists in the prostatectomy cohort, the MRUSuNet showed significantly higher specificity (0.88 vs 0.78; P = 0.004) and lesion Dice score (0.32 vs 0.27; P = 0.048), with no observed statistical difference in sensitivity (0.79; P = 1.000). CONCLUSION: Our findings demonstrate the potential of combining MRI and TRUS in a multimodal artificial intelligence (AI) model to improve localisation of clinically significant prostate cancer on TRUS images relative to single-modality models. These findings support further study of such multimodal AI models as an adjunct to MRI-TRUS fusion workflows.