Validation

Study Cohorts

MyoPath was validated on 478 H&E whole slide images from two independent cohorts:

Cohort	n	Source	Composition
GTEx	399	Genotype-Tissue Expression Project	Autopsy skeletal muscle; healthy controls and subclinical wasting spectrum
HuashanMuscle	79	Huashan Hospital, Fudan University	Clinical muscle biopsies; DM1, LGMD, inflammatory myopathy, and healthy controls

The GTEx cohort served as the training set; HuashanMuscle served as the independent external validation set. No retraining was performed on HuashanMuscle data.

Segmentation Performance

Segmentation accuracy was assessed by Dice coefficient and Intersection-over-Union (IoU) against manual expert annotations:

Tissue Layer	Method	Dice	IoU
Myofiber	Cellpose-SAM	0.92 ± 0.03	0.85 ± 0.06
Fat	Pixel classifier	0.95 ± 0.02	0.91 ± 0.03
Nucleus	Watershed	0.87 ± 0.04	0.78 ± 0.06
Connective tissue	Boolean subtraction	0.88 ± 0.04	0.78 ± 0.06

Reproducibility

Intraclass correlation coefficients (ICC) for all seven pathology indicators exceeded 0.88, indicating high inter-rater and intra-rater reproducibility of the automated pipeline.

MyoPath Score Performance

Discrimination

Dataset	AUC	Metric
GTEx (LOO-CV)	0.735	Leave-one-out cross-validation
GTEx (full model)	0.788	Training set AUC
HuashanMuscle	0.873	External validation (no retraining)

The higher AUC on the external validation set suggests that clinically diagnosed myopathy (HuashanMuscle) produces more separable morphometric changes than subclinical wasting (GTEx).

Primary Biomarkers

Two indicators emerged as the strongest individual discriminators:

Biomarker	p-value	Effect size (rank-biserial r)
NCI	$1.3\times {10}^{-5}$	0.69
Fiber CV	$2.9\times {10}^{-4}$	0.58

Both carry the largest standardized regression coefficients in the MyoPath Score model (${\beta }_{\text{NCI}}=0.57$, ${\beta }_{\text{CV}}=0.53$).

Clinical Correlations

Disease-Specific Findings

Myotonic dystrophy type 1 (DM1):

• NCI was highest in DM1 (median 0.121), consistent with centronuclear pathology
• NCI correlated with CTG repeat count: Spearman $\rho =0.46$, $p=0.042$
• Fiber CV inversely correlated with grip strength: $\rho =-0.61$, $p=0.031$

Limb-girdle muscular dystrophy (LGMD):

• Fiber CV showed a dose-response with mutation severity:
  • 2× Missense: CV = 0.44
  • LoF + Missense: CV = 0.49
  • 2× LoF: CV = 0.65

GTEx wasting spectrum:

• NCI showed a significant dose-response trend across wasting severity grades (Jonckheere-Terpstra $p<{10}^{-4}$)

Cross-Modal Validation with MyoScore

MyoPath pathology metrics correlate with MyoScore's transcriptomic dimensions, demonstrating cross-modal consistency:

MyoScore Dimension	MyoPath Indicator	Direction
LeanMuscle	Fat infiltration	Negative correlation
Youth	Fibrosis extent	Negative correlation
GMHS composite	Fiber CV	Negative correlation

This convergence between transcriptomic and morphometric assessments supports the biological validity of both tools.

Limitations

• ROI dependency: Tissue composition metrics (fat, fibrosis) vary with ROI placement. NCI and Fiber CV are more robust.
• H&E only: The pipeline is optimized for H&E stains. Special stains (ATPase, NADH) and immunohistochemistry are not supported.
• Single ROI: Current validation used one ROI per slide. Multi-ROI strategies may improve representativeness for heterogeneous biopsies.
• Connective tissue estimation: Boolean subtraction inherits errors from muscle and fat segmentation layers.
• Population scope: Training was performed on GTEx (predominantly European ancestry). Performance on other populations requires further validation.