Hip prosthesis artifacts degrade deep-learning prostate MRI diagnosis

In prostate MRI, moderate-to-severe hip prosthesis artifacts degraded all three deep-learning models (AUC 0.62–0.71) compared with exams without prostheses (AUC 0.74–0.81), while radiologist PI-RADS performance held up better (AUC 0.77–0.79 across artifact categories).

• Three convolutional-neural-network classifiers (T2WI alone; T2WI+DWI+ADC; T2WI+DCE) were trained on artifact-free exams and tested on 416 exams with hip prostheses and 2,080 matched controls.
• For Gleason ≥7 cancer, all DL models dropped below radiologist performance when moderate-to-severe artifact was present; no model surpassed PI-RADS in that setting.
• This was a single-center retrospective design and the models were not retrained or fine-tuned on artifact-containing exams, which the authors note as a limitation.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.