Machine learning-assisted SHG morphometry reveals distinct collagen microarchitectures of trabecular bone and fibrosis in bone marrow biopsies

Technical Summary: Machine Learning-Assisted SHG Morphometry of Bone Marrow Collagen

Problem Statement
Collagen microarchitecture in bone marrow biopsies remains a largely underexplored source of quantitative biomarkers for histopathological diagnostics. While Second Harmonic Generation (SHG) microscopy is increasingly used to assess fibrosis, the specific collagen organization of trabecular bone within trephine biopsies is poorly characterized. This gap is critical because myeloproliferative neoplasms, such as primary myelofibrosis, involve not only fibrotic collagen accumulation but also profound alterations in bone remodeling and osteosclerotic microarchitecture. Furthermore, existing SHG analysis often lacks a standardized, reproducible workflow for segmenting complex fibrillar structures, and the influence of annotation strategies on downstream morphometric measurements is not fully understood.

Methodology
The study employed a combination of high-resolution SHG microscopy and shallow machine learning-assisted morphometry to analyze bone marrow trephine biopsies from three patients diagnosed with primary myelofibrosis.

• Imaging: SHG imaging was performed on a Leica TCS SP5MP confocal microscope using backward detection. Due to photodamage limitations with high-NA objectives, a 10×/0.4 objective was used to achieve a pixel size of 200 nm, generating 4096 × 4096 pixel images.
• Segmentation (LabKit): The authors utilized the LabKit plugin within Fiji to create shallow learning classifiers for segmenting collagen fibers. Six distinct annotation strategies were evaluated, varying in foreground sampling (SHG intensity, fiber straightness, fiber number) and background placement.
• Classifier Optimization: The study systematically compared classifiers to determine which annotation strategy best preserved fibrillar structures across varying signal intensities. Based on this comparison, a set of "training rules" for efficient LabKit classifiers was formulated and validated using Dice coefficients and Intersection over Union (IoU) metrics against expert-generated masks.
• Workflow Development: Two Fiji-based pipelines were developed for semi-automated and automated analysis. These pipelines integrated the optimal LabKit classifier with morphometric plugins: "Local Thickness" for fiber width analysis and "Skeleton" for network geometry (branching, branch length).
• Quantitative Analysis: Ten 80×80 µm Regions of Interest (ROIs) were selected for each tissue type (structured trabecular bone, unstructured trabecular bone, and fibrosis) per patient. Quantitative parameters included fiber width distribution, network branching complexity, branch length, and angular dispersion (using the Fiji Directionality plugin). Statistical analysis relied on estimation statistics (mean paired differences with bootstrap confidence intervals) rather than null-hypothesis significance testing.

Key Contributions

1. Distinct Microarchitectures: The study establishes that structured trabecular bone, unstructured trabecular bone, and fibrosis represent three distinct collagen microarchitectural states, characterized by specific differences in fiber thickness, branching complexity, and orientation.
2. Annotation Guidelines: The paper formulates and experimentally validates specific "rules of thumb" for training shallow learning classifiers on SHG images. It demonstrates that annotation strategy (e.g., placing background lines on both sides of the foreground, spanning the full fibril width) critically influences segmentation accuracy and the preservation of low-intensity fibers.
3. Reproducible Workflow: The authors provide a validated, affordable, and reproducible workflow integrating LabKit shallow learning with Fiji morphometry, capable of processing large images (4096 × 4096) with an 8.5-fold increase in analysis speed compared to manual processing.
4. Differentiation of Tissue Types: The study successfully differentiates between fibrosis and bone remodeling states using quantitative descriptors, showing that fibrosis exhibits thinner fibers with reduced branching, while unstructured bone shows extensive branching with shorter skeleton segments.

Results

• Classifier Performance: Classifier 4, trained using specific foreground/background proximity rules and continuous fiber annotation, demonstrated the best overall performance. It accurately segmented structures across the full SHG intensity range, avoided artificial merging of bright structures, and preserved low-intensity fibers that other classifiers missed.
• Fiber Thickness: Structured trabecular bone showed a consistent fiber thickness distribution with a peak at approximately 4 µm. In contrast, both fibrosis and unstructured bone exhibited thinner fibers and greater variability in thickness.
• Network Geometry:
  • Structured Bone: Exhibited a highly ordered network with parallel fiber orientation and lower angular dispersion.
  • Unstructured Bone: Showed extensive network branching and systematically shorter branch lengths compared to structured bone, consistent with remodeling-associated alterations.
  • Fibrosis: Characterized by significantly thinner collagen fibers and reduced branching complexity compared to structured bone.
• Orientation: Structured trabecular bone displayed pronounced preferred fiber orientations, whereas unstructured bone and fibrosis showed broader, more isotropic orientation distributions.

Significance and Claims
The authors claim that this study establishes a novel framework for the quantitative assessment of bone marrow pathology that goes beyond conventional collagen abundance grading. By providing quantitative descriptors of collagen organization, branching architecture, and structural heterogeneity, the approach offers a new dimension for diagnostics and mechanistic studies of bone marrow remodeling.

The paper emphasizes that shallow learning approaches, particularly when guided by the proposed annotation rules, offer practical advantages for exploratory histopathology, including rapid classifier generation and direct integration into standard workflows like Fiji. The authors assert that their method provides "explainable" morphometric descriptors (e.g., fiber thickness, branch length) that are directly linked to observable biological features, facilitating validation by pathologists. While the study focuses on primary myelofibrosis, the authors suggest the workflow is applicable to a broad range of fibrosis- and bone-related pathologies, including osteoporosis and tumor stroma analysis, without claiming immediate clinical deployment or correlation with specific WHO fibrosis grades in this specific dataset.