Cross-Cohort Automatic Knee MRI Segmentation With Multi-Planar U-Nets.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in radiology
• Knee osteoarthritis research

Background:

• Manual segmentation of medical image volumes is time-consuming.
• Automatic segmentation tools often lack generalizability across different clinical settings and patient cohorts.

Purpose of the Study:

• To evaluate the performance of the open-source Multi-Planar U-Net (MPUnet) against the Knee Imaging Quantification (KIQ) framework and a 2D U-Net.
• Assess algorithm behavior on three distinct clinical cohorts without extensive adaptation.

Main Methods:

• Retrospective study of 253 subjects from three knee osteoarthritis (OA) studies (CCBR, OAI, PROOF).
• Evaluation of MPUnet, KIQ, and 2D U-Net on knee MRI scans with varying demographics and OA severity (KL grades 0-4).
• Models were assessed without tuning using Dice coefficients and statistical tests (Wilcoxon signed-rank).

Main Results:

• MPUnet demonstrated superior or equal performance to KIQ and 2D U-Net across all knee compartments and cohorts.
• Mean Dice overlap for MPUnet was significantly higher in CCBR and OAI cohorts compared to KIQ and 2D U-Net.
• MPUnet showed improved performance on KL grade 3 scans in the CCBR cohort.

Conclusions:

• MPUnet matches or surpasses state-of-the-art knee MRI segmentation models across diverse cohorts, sequences, and patient demographics.
• The MPUnet's lack of need for manual tuning makes it both accurate and user-friendly for clinical application.