Improving Robustness of Deep Learning Based Knee MRI Segmentation: Mixup and Adversarial Domain Adaptation

TL;DR

This paper tackles robust knee cartilage segmentation across heterogeneous knee MRI acquisitions by evaluating two regularization strategies: Mixup and adversarial Unsupervised Domain Adaptation (UDA). Using a strong 2D U-Net baseline and three knee DESS MRI datasets from different scanners, the authors quantify cross-domain generalization and OA-severity effects. They find that Mixup consistently improves robustness with lower computational cost, while UDA also helps but can degrade source-domain performance and requires heavier training. The work demonstrates practical improvements in clinically relevant knee regions and provides public release of code and pre-trained models to support broader adoption in OA imaging and drug-development workflows.

Abstract

Degeneration of articular cartilage (AC) is actively studied in knee osteoarthritis (OA) research via magnetic resonance imaging (MRI). Segmentation of AC tissues from MRI data is an essential step in quantification of their damage. Deep learning (DL) based methods have shown potential in this realm and are the current state-of-the-art, however, their robustness to heterogeneity of MRI acquisition settings remains an open problem. In this study, we investigated two modern regularization techniques -- mixup and adversarial unsupervised domain adaptation (UDA) -- to improve the robustness of DL-based knee cartilage segmentation to new MRI acquisition settings. Our validation setup included two datasets produced by different MRI scanners and using distinct data acquisition protocols. We assessed the robustness of automatic segmentation by comparing mixup and UDA approaches to a strong baseline method at different OA severity stages and, additionally, in relation to anatomical locations. Our results showed that for moderate changes in knee MRI data acquisition settings both approaches may provide notable improvements in the robustness, which are consistent for all stages of the disease and affect the clinically important areas of the knee joint. However, mixup may be considered as a recommended approach, since it is more computationally efficient and does not require additional data from the target acquisition setup.

Figures (5)

• Figure 1: Examples of MRI images from Datasets A, B, and C. Here, we show only the tibiofemoral areas, which enclose femoral cartilage, tibial cartilage, and menisci.
• Figure 2: Schematic view of our approaches -- without \ref{['fig:scheme_bl']} and with \ref{['fig:scheme_da']} UDA. Mixup, if used, is applied only during the training. In UDA setting \ref{['fig:scheme_da']}, the S- and D- networks are trained in an adversarial manner. During the testing, only the S-network is utilized.
• Figure 3: DESS MRI scan \ref{['fig:mri_scan']} and the annotations of knee cartilage and meniscal tissues \ref{['fig:mri_mask']}, both rescaled to isotropic resolution. White lines in \ref{['fig:mri_scan']} indicate the orientation of sagittal slices.
• Figure 4: Distributions of the planar DSCs computed slice-wise (from 0th to 159th slice, medial to lateral, respectively). Solid lines indicate the distribution means, bright bands -- the 95% confidence intervals. Slices approx. 20-60 and 100-140 correspond to the locations of the medial and lateral femoral condyles (i.e. weight-bearing areas of the joint). Slices approx. 60-100 enclose the intercondylar notch and, therefore, are of less clinical interest.
• Figure 5: Example images of tibiofemoral contact zones from Datasets A and C, respective annotations, and the segmentation masks produced by the baseline and the regularized approaches. Visual differences between the datasets can be observed. Colors highlight cartilage tissues: orange -- femoral, yellow -- tibial, purple -- menisci. Patellar cartilage was not presented in the considered Dataset A slice, for that reason patellofemoral zone is not shown.