LM-CartSeg: Automated Knee MRI Segmentation

• LM-CartSeg is an automatic pipeline for segmenting knee cartilage and subchondral bone using deep learning and geometric rules to achieve anatomically meaningful lateral and medial compartmentalization.
• It integrates dual 3D nnU-Net models with zero-shot fusion and connected-component filtering, significantly improving segmentation accuracy metrics like DSC and ASSD.
• The system enables high-throughput OA radiomics analysis by extracting standardized features from 10 ROIs per knee along with automated quality control for clinical and multi-centre research.

LM-CartSeg is a fully automatic, quality-controlled pipeline for the segmentation of cartilage and subchondral bone in knee MRI, with explicit geometric compartmentalization into lateral/medial (L/M) regions and a comprehensive radiomics backend supporting high-throughput region-of-interest (ROI) analysis. LM-CartSeg is designed to address the need for robust, anatomically meaningful ROIs that jointly capture cartilage and subchondral bone in the context of radiomics studies of osteoarthritis (OA). The system integrates state-of-the-art deep learning models, deterministic geometric rules for sub-compartmentation, and thorough feature extraction and quality control protocols, enabling efficient processing for multi-centre OA research and large-scale imaging studies (Zhang, 3 Dec 2025).

1. Pipeline Overview and Rationale

LM-CartSeg combines two independently trained 3D nnU-Net models, a geometric approach to subchondral band construction, a data-driven L/M split, and automated radiomics feature extraction. The pipeline’s components are summarized as follows:

• Dual 3D nnU-Net segmentation models trained on SKM-TEA and OAIZIB-CM datasets, each with distinct anatomical labeling protocols.
• Zero-shot model fusion at test time by voxel-wise ensemble averaging, followed by connected-component filtering to remove implausible segmentation artifacts.
• Geometric construction of subchondral bone bands of fixed physical thickness beneath the cartilage–bone interface.
• Robust, deterministic partitioning of cartilage and bands into lateral and medial compartments using PCA and k-means clustering.
• Comprehensive radiomics feature extraction from ten standardized ROIs per knee, including both cartilage and subchondral bands for each compartment.
• Automated anatomical quality control (QC) via volume and thickness signatures, with flagging of outlier cases.

This approach alleviates reliance on manual ROI definition and template-based compartmentalization, while supporting high-throughput multi-centre studies and minimizing domain shift effects across cohorts and image acquisition protocols.

2. Deep Learning Segmentation and Post-Processing

Two 3D nnU-Net models, following Isensee et al. (2021, 2024), are trained from scratch on SKM-TEA (138 knees) and OAIZIB-CM (404 knees) MRI datasets. The models use nnU-Net’s default preprocessing (resampling to dataset-specific voxel spacings, $z$-score normalization) and are optimized using a composite loss: $$\mathcal{L} = \mathcal{L}_{\mathrm{CE}} + \mathcal{L}_{\mathrm{Dice}},$$ where

$$\mathcal{L}_{\mathrm{CE}} = -\sum_{x\in V}\sum_{\ell=1}^L G_\ell(x)\,\log P_\ell(x)$$

and

$$\mathcal{L}_{\mathrm{Dice}} = 1 - \frac{2\sum_x G_\ell(x)\,P_\ell(x)}{\sum_x G_\ell(x)+\sum_x P_\ell(x)}.$$

During inference, both models generate probability maps on each input, which are fused by voxel-wise averaging: $$P_{\mathrm{fuse}}(x) = \tfrac12 \left(P^{(1)}(x) + P^{(2)}(x)\right),$$ with @@@@1@@@@ label assignment via $$\hat{\ell}(x) = \arg\max_\ell\,P_{\mathrm{fuse},\ell}(x)$$.

Connected-component analysis is applied to binary masks per label:

• For bones, only the component closest to the image center is retained (or those above a minimum volume threshold, discarding components touching image borders).
• For cartilage, all components of size at least $n_{\min} = \lfloor V_{\min}/(s_x s_y s_z)\rfloor$ are kept, with $V_{\min}$ a label-specific minimum volume in mm³ and $(s_x, s_y, s_z)$ the dataset-specific voxel spacings.

This fusion strategy mitigates domain-specific failure modes and, together with post-processing, sharpens surface metrics (macro-averaged ASSD improved from 2.63 mm to 0.36 mm, HD95 from 25.16 mm to 3.35 mm, DSC from 0.892 to 0.907 on OAIZIB-CM; $p<10^{-9}$).

3. Geometric Subchondral Band Extraction and Lateral/Medial Split

Subchondral bone bands are constructed using a distance-based shell beneath the cartilage–bone interface: $$\Omega_{\mathrm{sub}} = \{\,x \mid B(x)=1 \wedge D(x)\leq t\,\},$$ where $B(x)$ is the binary bone mask, $C(x)$ is the cartilage mask, $D(x)$ is the Euclidean distance from the nearest cartilage voxel in physical space, and $t=10$ mm defines the band thickness.

For data-driven lateral/medial compartmentalization, the procedure is as follows:

• The cartilage or fallback subchondral mask is sampled as a point cloud $X=\{x_i\}$ in world coordinates.
• The centroid $c$ and covariance $\Sigma$ are computed to establish eigenvector axes. The superior–inferior (SI) axis is extracted as $s = e_3/\|e_3\|$, and in-plane coordinates are defined post SI-projection via a second 2D PCA on $x_i^\perp$.
• The resulting planar point set $z_i$ is clustered into $k=2$ compartments by standard $k$-means, with cluster centroids mapped back to 3D. Lateral/medial assignment is determined via the scanner-space $x$ coordinate and established laterality (left versus right knee).
• The k-means model is then applied for per-voxel compartment assignment to all cartilage and subchondral regions.

This deterministic geometric rule demonstrated domain-general stability (no observed "side swaps" under domain shift), unlike direct nnU-Net-based L/M segmentation, which was domain sensitive (Zhang, 3 Dec 2025).

4. Radiomics Feature Extraction and Quality Control

Radiomics extraction spans 10 ROIs per knee (LF, MF, LT, MT, LP, MP, and subchondral bands), resulting in up to 4,650 non-shape PyRadiomics features per case. These features include:

• First-order histogram statistics (mean, skewness, kurtosis, entropy, etc.).
• Texture features (GLCM, GLRLM, GLSZM, NGTDM, GLDM) computed on discretized images (fixed bin width = 5).
• Filtered image features, using LoG filtering and multi-scale wavelet transforms.

Automated anatomical QC leverages both volume and mean thickness ($\bar T_R = 2V_R/A_R$ with $A_R$ from marching cubes), in conjunction with percentile statistics of intra-ROI distance transforms. Lateral-medial asymmetry is tracked using ratios (e.g., tibial thickness ratio $$r_{\mathrm{TIB}} = \bar T_{\mathrm{MT}}/\bar T_{\mathrm{LT}}$$) and differences. Outlier detection in $V_R$, $\bar T_R$, or $r_{\mathrm{TIB}}$ prompts review.

5. Metrics and Performance Evaluation

Segmentation accuracy is assessed using multiple surface and overlap metrics:

• Dice Similarity Coefficient (DSC): $\frac{2|G \cap P|}{|G| + |P|}$.
• Average Symmetric Surface Distance (ASSD): mean bidirectional surface distance.
• 95th-percentile Hausdorff Distance (HD95): robustness to outlier boundary points.

Quantitative results demonstrate macro-averaged ASSD, HD95, and DSC improvements post-processing. On OAIZIB-CM, ASSD improved from 2.63 mm to 0.36 mm, HD95 from 25.16 mm to 3.35 mm, while DSC rose from 0.892 to 0.907. In zero-shot SKI-10, DSC reached 0.797, ASSD dropped to 0.94 mm, and HD95 fell to 6.13 mm.

6. Radiomics Feature Analysis and Osteoarthritis Discrimination

Spearman correlations between individual radiomics features and corresponding ROI volume and thickness revealed that only 6–12% of features per ROI are strongly size-dependent ($|\rho| \geq 0.7$), with lower dependence for subchondral bands than for cartilage.

To test OA (KL≥2) vs. non-OA (KL≤1) discrimination, four classifier families (LASSO logistic regression, $k$NN, RBF-SVM, XGBoost) were trained in 5-fold cross-validation on:

• A set of LASSO-selected radiomic features weakly correlated with size (primarily texture).
• A set restricted to features strongly size-linked.

On OAIZIB-CM, radiomics classifiers (texture dominated) achieved AUCs of 0.85–0.91, whereas those restricted to size-linked features yielded AUCs of 0.73–0.78 (approx. 0.12 absolute decrease). This pattern recurred in an independent cohort, indicating that textural features beyond morphometry are essential for discriminative OA classification (Zhang, 3 Dec 2025).

7. Summary of Contributions and Applications

Key contributions of LM-CartSeg include:

• A plug-and-play deep-learning and geometric rules pipeline producing anatomically standardized L/M cartilage and bone ROIs without manual annotation.
• Zero-shot nnU-Net fusion and connected-component filtering that enhances surface metrics while preserving region overlap.
• A deterministic PCA + k-means L/M split resilient to scanner, cohort, and OA severity differences.
• Automated QC based on volume and thickness signatures with explicit flagging of anatomical outliers.
• A large-scale radiomics feature analysis demonstrating that texture features offer OA discrimination advantages beyond cartilage or bone morphometry.

LM-CartSeg supports large-scale, multi-centre knee OA radiomics studies, longitudinal disease progression modeling, and clinical deployment in MRI workflows for early OA detection and risk stratification, with a fully automatic, quality-controlled, and reproducible approach (Zhang, 3 Dec 2025).