Methods

Segmentation Pipeline Architecture

MyoPath implements a four-layer tissue segmentation pipeline on H&E-stained skeletal muscle whole slide images (WSI). Each layer uses a method suited to its biological target.

WSI (H&E) → ROI Selection (1500 µm²)
                 │
                 ├─→ Layer 1: Cellpose-SAM → Myofiber instances
                 ├─→ Layer 2: Pixel classifier → Fat regions
                 ├─→ Layer 3: Watershed → Nuclei
                 └─→ Layer 4: Boolean subtraction → Connective tissue
                 │
                 ▼
         37 morphometric features → 7 pathology indicators → MyoPath Score

Layer 1: Myofiber Instance Segmentation

Method: Cellpose-SAM (Cellpose with Segment Anything Model)

Cellpose is a generalist deep learning model for cellular segmentation that uses gradient flow tracking. In MyoPath, the Cellpose model segments individual myofibers as polygon instances rather than semantic masks, enabling per-fiber morphometric analysis.

Key parameters:

• downsample: 10.0 — controls tile size to prevent memory overflow
• cellprobThreshold: 0 — probability threshold for cell detection
• GPU acceleration recommended (~30–60 s per ROI)

Output: Individual myofiber polygons with cross-sectional area, perimeter, shape factor, and aspect ratio computed per fiber.

Layer 2: Fat Infiltration Detection

Method: Pixel classifier (trained in QuPath)

A random-forest pixel classifier named "fat in muscle" is trained on representative H&E patches to distinguish adipose tissue from background and other tissue types. The classifier operates at the pixel level, producing smooth fat region boundaries.

Training:

• Annotate fat vs. non-fat regions in 10–20 representative slides
• Train via Classify → Pixel classifiers → Train pixel classifier in QuPath
• Save the classifier as "fat in muscle" for pipeline compatibility

Output: Fat region polygons with area measurements. Fat-muscle overlapping regions are excluded.

Layer 3: Nuclear Detection

Method: Watershed segmentation

A watershed algorithm detects individual nuclei within the ROI. Detected nuclei are then spatially assigned to tissue compartments (muscle, connective, or unassigned) based on their centroid location relative to myofiber and fat polygons.

Nuclear classification:

• Muscle nuclei — centroid falls within a myofiber polygon
• Connective nuclei — centroid falls within the connective tissue region
• Unassigned nuclei — centroid falls within fat or on annotation boundaries

Nuclear localization: For each muscle nucleus, a normalized radial position ${\rho }_k$ is computed (see NCI definition), classifying nuclei into peripheral (${\rho }_k\le 0.3$), intermediate ($0.3<{\rho }_k<0.7$), or central (${\rho }_k\ge 0.7$) zones.

Layer 4: Connective Tissue Estimation

Method: Boolean subtraction

Connective tissue is not directly segmented. Instead, it is estimated as the residual area after subtracting muscle and fat from the ROI:

$$A_{\text{connective}}=A_{\text{ROI}}-A_{\text{muscle}}-A_{\text{fat}}$$

This approach avoids the difficulty of training a dedicated connective tissue classifier on H&E stains, where endomysial and perimysial collagen has variable appearance. The trade-off is that any segmentation errors in muscle or fat propagate into the connective tissue estimate.

Feature Extraction

From the four tissue layers, MyoPath computes 37 morphometric features organized into five categories:

Category	Features	Key indicators
Tissue composition	10	Areas and percentages of muscle, fat, connective tissue, nuclei
Fiber size	7	Mean, median, std, min, max, Q1, Q3 of fiber cross-sectional area
Fiber shape	3	Shape factor (circularity), aspect ratio and their variability
Nuclear distribution	7	Nuclei counts per compartment, nuclei per fiber statistics
Nuclear localization	3	Peripheral ratio, central ratio, multinucleated fiber count

See Morphometric Features for the complete feature catalog.

MyoPath Score

Seven clinically interpretable pathology indicators are derived from the 37 features. These are combined via logistic regression into a composite MyoPath Score:

$$\text{MyoPath Score}=\frac{1}{1+e^{-z}}$$$$z={\beta }_0+{\beta }_1\cdot \text{NCI}+{\beta }_2\cdot \text{Fiber CV}+{\beta }_3\cdot \text{Shape}+{\beta }_4\cdot \text{Fat\%}+{\beta }_5\cdot \text{Fibrosis\%}+{\beta }_6\cdot \text{NMR}+{\beta }_7\cdot \text{Inflammation}$$

The model was trained on the GTEx cohort (n = 399) with leave-one-out cross-validation and validated on the independent HuashanMuscle cohort (n = 79) without retraining.

INFO

NCI and Fiber CV carry the largest standardized regression coefficients ($\beta =0.57$ and $0.53$), confirming their role as primary biomarkers.

ROI Selection Strategy

MyoPath uses a standardized 1500 µm × 1500 µm ROI to ensure:

• Sufficient sampling: ~200–600 myofibers per ROI (median ~400)
• Computational feasibility: single-ROI processing in < 2 min
• Cross-study comparability: fixed area eliminates scaling confounds

The ROI is positioned manually by the operator at the region of interest. For clinical validation, we recommend analyzing at least one ROI per biopsy section, positioned over the most representative (or most affected) area.

WARNING

Tissue composition metrics (fat infiltration, fibrosis) are inherently ROI-dependent. NCI and Fiber CV are more robust to ROI placement because they are computed per-fiber and normalized.