Deep learning-based prostate cancer diagnosis on MRI with hip prostheses: artifact and sequence effects.

objectiveTo evaluate the impact of hip prosthesis-induced artifacts on the diagnostic performance of deep learning (DL)-based prostate cancer diagnosis on MRI and to investigate the optimal MRI sequence combination for DL analysis.

methodsThis retrospective study included prostate MRI examinations performed between 2017 and 2023. Three DL-based image classification models were developed using examinations from patients without hip prostheses, with different input combinations: T2-weighted imaging (T2WI) alone, T2WI plus diffusion-weighted imaging (DWI) and apparent diffusion coefficient maps, and T2WI plus dynamic contrast-enhanced MRI. Test sets consisted of patients with hip prostheses who underwent prostate biopsy within one year after MRI and were stratified by artifact severity (mild, moderate, severe), as well as a matched test set of patients without hip prostheses. Diagnostic performance for Gleason score ≥ 7 prostate cancer was assessed using the area under the receiver operating characteristic curve (AUC) and compared with PI-RADS assessments.

resultsThe test sets included 416 examinations with and 2,080 matched examinations without hip prostheses. All DL models showed reduced diagnostic performance in the presence of moderate-to-severe susceptibility artifacts compared with examinations without prostheses (AUC range, 0.62-0.71 vs. 0.74-0.81, respectively). Across moderate-to-severe artifact categories, PI-RADS outperformed all DL models (AUC range, 0.77-0.79).

conclusionDL-based models exhibited limited robustness to hip prosthesis-induced susceptibility artifacts, whereas radiologist performance remained relatively preserved. These findings highlight current limitations of DL-based prostate cancer diagnosis in patients with hip prostheses and underscore the continued importance of expert radiologist interpretation.