ERANet: Edge replacement augmentation for semi-supervised meniscus segmentation with prototype consistency alignment and conditional self-training

TL;DR

ERANet is a novel semi-supervised framework designed to segment the meniscus from MRI scans using minimal labeled data. By combining a domain-specific "Edge Replacement" augmentation with feature-space alignment and a stability-based self-training strategy, it significantly outperforms existing methods, especially in challenging, low-contrast imaging sequences.

What Problem Does This Paper Address?

The meniscus—the crescent-shaped cartilage in your knee—is a primary indicator of osteoarthritis. However, segmenting it from MRI scans is a nightmare for computer vision models. It is small, its shape varies wildly between patients, and it often blends into surrounding tissues (low contrast), particularly in Fast Spin Echo (FSE) sequences where the meniscus appears almost entirely dark.

While deep learning has made strides here, it usually requires massive amounts of hand-labeled data, which is expensive and time-consuming for radiologists to produce. Semi-supervised learning (SSL) aims to fix this by using a few labeled images alongside a mountain of unlabeled ones. The problem? Standard SSL "tricks" (like generic rotations or color jitters) don't respect the unique anatomy of the knee, often leading to "noisy" labels that confuse the model rather than help it.

Methods

The authors introduce ERANet, which operates within a "Mean Teacher" architecture (where a student model learns from a teacher model that is a moving average of itself). The framework introduces three core innovations:

1. Edge Replacement Augmentation (ERA): Unlike standard "Copy-Paste" augmentations that move patches between different images, ERA is an intra-image strategy. It identifies the extreme points of the meniscus and replaces the peripheral regions with background context from the same slice. This simulates the "edge erosion" and morphological variability seen in real-world clinical cases, forcing the model to become robust at finding indistinct boundaries.
2. Prototype Consistency Alignment (PCA): This module works in the feature space. It calculates "prototypes"—essentially the average feature representation for a class (meniscus vs. background). It then enforces "feature compactness," ensuring that unlabeled pixels are mathematically similar to the meniscus prototype, which helps the model distinguish the tissue even when the visual contrast is weak.
3. Conditional Self-Training (CST): Self-training usually involves a model generating "pseudo-labels" for unlabeled data and then training on them. To avoid "confirmation bias" (where the model learns from its own mistakes), CST only accepts pseudo-labels that are temporally stable. If the model's prediction for a specific unlabeled pixel stays consistent across five different checkpoints during training, it’s deemed reliable enough to be used as ground truth.

Datasets

The researchers validated ERANet on two distinct MRI sequences to test its generalization:

• OAI-iMorphics: A public dataset of 176 examinations using 3D Double Echo Steady State (DESS) sequences.
• In-house Dataset: 33 scans from patients with varying degrees of osteoarthritis, acquired using 3D Fast/Turbo Spin Echo (FSE/TSE) sequences. This dataset is particularly challenging because the meniscus appears hypointense (very dark) against a dark background.

Key Insights and Ideas

• Anatomy-Aware Perturbation Matters: Generic data augmentation can actually hurt medical models by creating "impossible" anatomy. ERA works because it simulates plausible pathological variations (like inflammation or meniscus thinning) that the model is likely to encounter in a clinic.
• Stability as a Proxy for Truth: In the absence of a "ground truth" label for an image, the authors realized that consistency over time is a powerful signal. If a model keeps saying "this is meniscus" throughout the training process, it is much more likely to be right than a model that flip-flops.
• Synergy is Key: The ablation studies showed that while ERA, PCA, and CST each help, their combination provides a "greater than the sum of its parts" boost. ERA provides diversity, PCA provides stability, and CST provides a clean signal for the final training stages.

Key Results

ERANet set a new state-of-the-art benchmark for semi-supervised meniscus segmentation:

• Minimal Data Performance: On the OAI dataset, using only 10% labeled data (14 scans), ERANet achieved a Mean Dice Similarity Coefficient (DSC) of 86.66%, nearly matching the performance of a fully supervised model trained on ten times as much data.
• Superiority over SOTA: It outperformed other popular semi-supervised frameworks like UA-MT and AdvChain across both datasets.
• Hard Case Success: On the difficult in-house FSE dataset, ERANet achieved a relative improvement of 27.7% over the baseline model when only 3 labeled scans were available, proving its utility in data-scarce clinical environments.

Limitations and Future Work

The authors acknowledge a few hurdles:

• Domain Specificity: The ERA module is specifically tailored to the meniscus. While effective, it wouldn't work out-of-the-box for other organs (like the liver or heart) without redesigning the "extreme point" logic.
• Complexity: Adding PCA and CST increases the computational overhead and training time compared to simpler architectures.
• Future Directions: The team plans to explore "adaptive reliability metrics" to replace fixed thresholds and investigate multi-modality fusion (combining different types of MRI) to further improve accuracy.

@misc{li2025eranet,
  author = {Siyue Li and Yongcheng Yao and Junru Zhong and Shutian Zhao and Fan Xiao and Michael Tim‐Yun Ong and Kevin Ki‐Wai Ho and James F. Griffith and Yudong Zhang and Shuihua Wang‎ and Jin Hong and Weitian Chen},
  title = {{ERANet: Edge replacement augmentation for semi-supervised meniscus segmentation with prototype consistency alignment and conditional self-training}},
  year = {2025},
  eprint = {2502.07331},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.07331},
  howpublished = {arXiv},
}