Impact of Scanner Manufacturer, Endorectal Coil Use, and Clinical Variables on Deep Learning-assisted Prostate Cancer...

Area of Science:

• Medical Imaging Informatics
• Artificial Intelligence in Oncology
• Radiomics and Machine Learning

Background:

• Accurate classification of prostate cancer (PCa) aggressiveness is crucial for treatment decisions.
• Deep learning (DL) models show promise for analyzing biparametric MRI (bpMRI) in PCa detection.
• The generalizability of DL models is often limited by variations in imaging hardware and protocols.

Purpose of the Study:

• To evaluate the influence of scanner manufacturer and scanning protocol on DL model performance for PCa aggressiveness classification using bpMRI.
• To assess the impact of data from different subgroups, including scanner manufacturers and endorectal coil (ERC) use, on model generalizability.
• To determine if clinical features improve DL model performance in PCa aggressiveness assessment.

Main Methods:

• Retrospective analysis of 5478 PCa bpMRI cases from the ProstateNet dataset across 13 centers.
• Development and testing of five DL models to predict PCa aggressiveness using minimal lesion information.
• Evaluation of model performance using area under the receiver operating characteristic curve (AUC), comparing performance across different scanner manufacturers (Siemens, Philips, GE) and ERC usage.

Main Results:

• The best DL model achieved an AUC of 0.73 when trained and tested on the full dataset.
• Models demonstrated higher performance when trained and tested on data from the same manufacturer (average AUC difference of 0.05, P < .001).
• Inclusion of clinical features (age, PSA, PI-RADS) did not significantly improve model performance (P = .24).

Conclusions:

• Scanner manufacturer and ERC use significantly affect the performance and feature distributions of DL models for automated PCa aggressiveness classification on bpMRI.
• Model generalizability is challenged by variations in imaging acquisition parameters.
• Future DL model development for PCa classification requires careful consideration of multi-center and multi-manufacturer data harmonization.