Deep Learning Prostate MRI Segmentation Accuracy and Robustness: A Systematic Review.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Prostate cancer diagnosis relies heavily on Magnetic Resonance Imaging (MRI) for accurate segmentation.
• Deep learning (DL) shows promise for automating prostate MRI segmentation, but its accuracy relative to expert radiologists needs systematic evaluation.
• Understanding the robustness of DL algorithms across diverse datasets and vendors is crucial for clinical adoption.

Approach:

• A systematic review was conducted by querying Embase, PubMed, Scopus, and Web of Science for studies on prostate MRI segmentation using DL.
• 48 English-language articles published up to July 31, 2022, met the inclusion criteria.
• Data extracted included DL algorithm performance, MRI vendor, and training dataset characteristics, with the Dice Similarity Coefficient (DSC) as the primary outcome measure.

Key Points:

• Most published DL algorithms (93%) achieved a DSC of 0.86 or higher for whole prostate gland segmentation, meeting expert level.
• Mean DSCs were 0.79 ± 0.06 for the peripheral zone, 0.87 ± 0.05 for the transition zone, and 0.90 ± 0.04 for the whole prostate gland.
• DL algorithms demonstrated high accuracy across major MRI vendors (GE, Philips, Siemens), with mean DSCs ranging from 0.91 to 0.92.

Conclusions:

• Deep learning algorithms demonstrate accuracy for prostate MRI segmentation that is similar to that of fellowship-trained diagnostic radiologists.
• The performance of DL algorithms is robust across various training data sizes, MRI vendors, and prostate zones.
• Future research should prioritize evaluating the robustness of DL segmentation and its impact on patient outcomes in real-world clinical settings.