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

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Prostate cancer diagnosis relies heavily on MRI.
• Deep learning (DL) models show promise for automating prostate cancer detection.
• Hip prostheses can cause artifacts in MRI, potentially impacting diagnostic accuracy.

Purpose of the Study:

• To assess the effect of hip prosthesis artifacts on DL-based prostate cancer diagnosis using MRI.
• To determine the optimal MRI sequence combination for DL analysis in the presence of artifacts.
• To compare DL model performance with radiologist (PI-RADS) assessments.

Main Methods:

• Retrospective analysis of prostate MRI scans (2017-2023).
• Development of three DL models using T2WI, DWI, ADC maps, and DCE-MRI.
• Evaluation of DL models on test sets with varying artifact severity (mild, moderate, severe) and matched controls without prostheses.

Main Results:

• Moderate-to-severe hip prosthesis artifacts reduced DL model diagnostic performance (AUC 0.62-0.71) compared to non-artifact scans (AUC 0.74-0.81).
• PI-RADS assessments by radiologists outperformed all DL models in moderate-to-severe artifact categories (AUC 0.77-0.79).
• DL models showed limited robustness to susceptibility artifacts.

Conclusions:

• Deep learning models are currently not robust to hip prosthesis-induced artifacts in prostate MRI.
• Radiologist interpretation remains crucial for accurate prostate cancer diagnosis in patients with hip prostheses.
• Further research is needed to improve DL model performance in artifact-affected MRI scans.