PaperlayerView exact revision on arXiv

Source-linked AI summary

MorphoBranch: A Fine-Structure-Preserving Workbench for Morphometric Analysis of Branched Cellular Structures

Song Zhiying, Ling Hanyi, Wu Junyi, Jiang Yangbo

arXiv:2610.00860v2eess.IVcs.CVq-bio.QM

TL;DR

Fine and weakly labeled processes can fragment or connect falsely, biasing skeleton-based measurements. MorphoBranch couples deterministic morphometry with optional human-reviewed natural-language refinement. Across neuronal axon and microglial datasets, it improved centerline preservation and length-estimation accuracy over evaluated baselines.

  • Problem

    Fine, weakly labeled, densely distributed processes and imaging artifacts obscure connectivity, while segmentation errors propagate into biased skeleton-based measurements and make analysis difficult to reproduce.

    Source: §1 · pp. 1–2

  • Method

    MorphoBranch combines deterministic image-to-graph morphometry that retains structural evidence for skeleton refinement with optional LLM-assisted plans reviewed before execution.

    Source: §1 · p. 2, §2.3 · p. 3

  • Results

    Across AxonMIP, AxonStack, and CellMorph, MorphoBranch improved centerline preservation and length-estimation accuracy over evaluated baselines, although Meijering filtering had slightly higher Dice and IoU on CellMorph.

    Source: §4 · p. 11, §5 · p. 11

  • Takeaways & Limitations

    The evaluated workflow supports fine-structure-preserving morphometric analysis with inspectable, reproducible refinement across neuronal axon and microglial fluorescence images.

    Source: §5 · p. 11

    The current focus on 2D images and projected structures cannot fully resolve branches overlapping at different depths; CellMorph has limited annotated images, and branch-level topology lacks direct exhaustive-annotation validation.

    Source: §4 · p. 11

Abstract

from arXiv · show

Background and Objectives: Fluorescence-labeled cellular arbors provide readouts of neuronal and microglial morphology, but fine and weakly labeled processes are prone to fragmentation and false connections that bias skeleton-based measurements. We present MorphoBranch, a fine-structure-preserving, human-reviewable workbench for morphometry of branched cellular structures. Methods: MorphoBranch combines a deterministic Morphometry Engine with an LLM-assisted Refinement Engine. The Mor- phometry Engine implements an image-to-graph workflow integrating multiscale structural evidence extraction, hysteresis segmen- tation, evidence-constrained skeleton refinement, and graph-based morphometry. The Refinement Engine maps natural-language requests to registered actions for parameter adjustment, preview execution, metric reporting, and unsupported-request handling, while image processing and quantitative computation remain deterministic and reviewable. Results: MorphoBranch was evaluated on two public neuronal axon datasets, AxonMIP and AxonStack, and the in-house Cell- Morph dataset of microglial fluorescence images. It achieved the highest Skeleton F1 and clDice and the lowest length-estimation error among the evaluated methods on all three datasets, while also achieving the highest Dice and IoU on AxonMIP and Axon- Stack. Across 150 natural-language tasks, the Refinement Engine achieved a 94.0% end-to-end success rate. Conclusions: These results demonstrate that MorphoBranch provides a reproducible, human-reviewable workflow for mor- phometric analysis of branched cellular structures. It supports fine-structure-preserving quantification across neuronal axon and microglial fluorescence images while maintaining inspectable and reproducible analysis workflows.

1. Introduction

Reliable morphometry of fine, weakly labeled, densely branched cellular structures remains difficult because segmentation errors propagate into skeleton-based measurements. MorphoBranch combines deterministic image analysis with human-reviewed workflow refinement to address these challenges.

1. Introduction

  • Thin, weakly labeled processes and imaging artifacts obscure connectivity, while missed processes, breaks, or false links propagate into biased length, branching, and radial measurements.Repeated ROI selection, parameter adjustment, and manual inspection also make analysis time-consuming, operator-dependent, and difficult to reproduce.Source: §1 · p. 1
  • Existing tools leave gaps in coordinated cross-stage topology control and measurement-level validation, as well as in traceable iterative refinement across interfaces and scripts.The paper motivates an integrated, human-reviewable system that addresses both limitations.Source: §1 · p. 2
  • MorphoBranch combines fine-structure-preserving extraction with topology-aware skeleton and graph analysis in a task-oriented workbench for fluorescence images.The workbench also provides data management, quality control, provenance tracking, and result export.Source: §1 · p. 2
  • The optional Refinement Engine turns natural-language requests into editable plans that users can preview, revise, and confirm before deterministic execution.Configurations and execution records are retained as provenance, separating flexible interaction from quantitative computation.Source: §1 · p. 2
  • System-level evaluations cover public neuronal axon datasets, the in-house CellMorph microglial dataset, and a 150-task natural-language interaction benchmark.The evaluations assess structural fidelity, morphometric accuracy, computational efficiency, and refinement reliability.Source: §1 · p. 2

2. Method

MorphoBranch organizes analysis around a deterministic Morphometry Engine, with optional human-reviewed refinement and supporting provenance-preserving workflow modules.

2.1. System Overview and Design Principles

  • The Morphometry Engine coordinates fine-structure-preserving extraction with topology-aware skeleton and graph analysis across processing stages.This design is intended to limit loss of faint processes and propagation of local artifacts into measurements.Source: §2.1 · p. 2
  • The Refinement Engine translates natural-language requests into configuration changes that users review and confirm before deterministic execution.The operational workflow separates flexible interaction from quantitative computation, which uses the same deterministic modules for confirmed adjustments.Source: §2.1 · p. 2
  • Supporting modules retain project data, ROIs, configurations, quality-control outputs, user decisions, and execution records so confirmed analyses can be inspected and replayed without the language model.This preserves provenance across the analysis workflow.Source: §2.1 · p. 2

2.2. Project and Workflow-State Management

MorphoBranch organizes project data and analysis artifacts into versioned workflow states, separating provisional previews from accepted results.

2.2. Project and Workflow-State Management

  • Each execution creates a versioned state whose artifacts are associated with their source image, configuration, and parent state.Projects link immutable ND2 or TIFF inputs and metadata to ROIs, spatial calibration, configurations, intermediate artifacts, quality-control outputs, and exported results.Source: §2.2 · p. 2
  • The workbench exports quantitative results, quality-control visualizations, refinement feedback, and reproducibility records for review and downstream analysis.Its architecture pairs a deterministic Morphometry Engine with an optional LLM-assisted Refinement Engine.Source: Figure 1 · p. 3
  • Preview states hold provisional artifacts that cannot overwrite raw inputs or confirmed outputs, while confirmed states represent the active accepted analysis.Configuration changes invalidate affected downstream artifacts while preserving compatible upstream results, supporting restoration, comparison, and deterministic reruns.Source: §2.2 · p. 3

2.3. Deterministic Morphometry Engine

The deterministic Morphometry Engine carries microscopy images through evidence-coupled extraction, topology refinement, and graph-based morphometric analysis. It retains intermediate representations for inspection, provenance tracking, and reproducible reruns.

2.3. Deterministic Morphometry Engine

  • The engine couples a structure mask with continuous image evidence so topology refinement can repair supported discontinuities and reject unsupported connections before computing morphometric descriptors.Intermediate outputs are retained for quality-control inspection, provenance tracking, and deterministic reruns.Source: §2.3 · p. 3

2.3.1. Data Preparation

  • For image stacks, users select the channel, ROI, and z-range; a configured projection produces the analysis image passed to Structure Extraction.Precomputed two-dimensional images can be used directly, and image sequences can be registered to reduce frame-to-frame displacement.Source: §2.3.1 · p. 3

2.3.2. Structure Extraction

  • Multiscale ridge evidence and seed-guided hysteresis produce a binary mask while preserving weak processes connected to high-confidence seeds.Gaussian denoising and broad-scale background subtraction reduce high-frequency noise and slowly varying illumination.Source: §2.3.2 · p. 4

2.3.3. Topology Refinement

  • The refined skeleton graph reconnects candidate endpoint interruptions only when proximity, directional consistency, and continuous image support all pass their criteria.The graph is built after limited morphological closing and one-pixel skeletonization; adjacent junction pixels are consolidated to avoid inflated junction counts.Source: §2.3.3 · p. 4

2.3.4. Morphometric Analysis

  • Morphometric Analysis reports geometry, topology, branch hierarchy, and spatial complexity, including graph connectivity, branch lengths, Sholl analysis, and fractal dimension.Branch-hierarchy and Sholl summaries use a reference point when available; length- and area-based descriptors use physical units when spatial calibration is available, otherwise pixel units.Source: §2.3.4 · pp. 4–5, Table 1 · p. 5

2.4. LLM-Assisted Refinement Engine

The Refinement Engine translates natural-language requests into editable, registered workflow actions using current project context. Proposed changes are previewed and reviewed before confirmed adjustments are run by deterministic modules and recorded for replay.

2.4.1. State-Aware Planning and Tool Routing

  • Current project state and interaction history inform editable action plans that specify operations, parameter changes, affected stages, and expected outputs.The plan also uses configurations, preview outputs, QC results, and metric summaries.Source: §2.4.1 · p. 5
  • Proposed actions are checked against registered operations, parameter constraints, inputs, dependencies, and read-only fields; ambiguous or unsupported requests are returned for clarification.Routine requests route to built-in deterministic tools, while compatible external adapters can be supported.Source: §2.4.1 · pp. 5–6

2.4.2. Preview-Based Multi-Turn Refinement

  • Validated actions run in an isolated preview that reports configuration differences, updated outputs, and relevant quality-control indicators without changing raw inputs or confirmed outputs.Users can revise requests, compare previews, restore settings, or confirm adjustments; follow-up instructions apply to the current preview state.Source: Figure 4 · p. 5, §2.4.2 · p. 6

2.4.3. Confirmation and Deterministic Execution

  • Only explicitly confirmed adjustments enter the formal workflow, where deterministic Morphometry Engine modules produce measurements and configurations, outputs, tool calls, and user decisions are retained for review and replay.This separates natural-language interaction from image processing and quantitative computation.Source: §2.4 · p. 5, §2.4.3 · p. 6

2.5. Quality Control, Provenance, and Result Export

MorphoBranch retains quality-control outputs across image, graph, and measurement stages and records analysis inputs and decisions to support review and reruns.

2.5. Quality Control, Provenance, and Result Export

  • Quality-control outputs span analysis images, structural evidence, masks, skeletons, graph and branch visualizations, Sholl profiles, and metric summaries.Indicators include foreground occupancy, fragmentation, repair extent, and measurement changes, but are not automatic judgments of biological correctness.Source: §2.5 · p. 6
  • Each run records inputs, ROI and acquisition settings, calibration, processing parameters, reference information, outputs, and any assistant-mediated requests and tool calls.Measurements, QC visualizations, configurations, and audit records are exported in structured files for inspection, comparison, and reruns.Source: §2.5 · p. 6

2.6. System Implementation and Deployment

MorphoBranch is a modular Python application with a local Streamlit interface and a deterministic analysis pipeline that can run independently of the LLM service.

2.6. System Implementation and Deployment

  • The Python workbench uses a local Streamlit interface and scientific libraries for image processing, skeleton and graph analysis, visualization, and batch processing.The deterministic pipeline can run independently of the LLM service.Source: §2.6 · p. 6
  • The LLM layer uses a LangChain Agent with the DeepSeek API and registered MorphoBranch actions; evaluation used deepseek-chat at temperature zero.The reported runtime tests used a MacBook Pro with an Apple M5 processor and 32 GB unified memory, and deployment retains project data and records locally.Source: §2.6 · p. 6

3. Results

MorphoBranch was evaluated on public axon datasets, an in-house microglial dataset, and a 150-task interaction benchmark. It led on centerline and length metrics across the image datasets, while ablations, runtime measurements, example analyses, and interaction tests characterize its workflow.

3.1. Datasets and Experimental Setup

  • The evaluation covered AxonMIP and AxonStack axon images, 11 CellMorph microglial images, and 150 natural-language tasks spanning five categories.CellMorph references were manually annotated because dense ground-truth annotations were unavailable; its length was reported in pixels because spatial calibration was unavailable.Source: §3.1 · pp. 6–7, Table 2 · p. 8

3.2. Workbench Interface Demonstration

  • The workbench interface coordinates project inputs and parameters, intermediate outputs and morphometric summaries, and LLM-assisted refinement in three panels.The local interface integrates deterministic analysis, QC inspection, export, and natural-language refinement.Source: Figure 5 · p. 7, §3.2 · p. 7

3.3.1. Evaluation of Segmentation and Skeleton Preservation

  • MorphoBranch achieved the highest Skeleton F1 and clDice and lowest length-estimation error on all three datasets, plus the highest Dice and IoU on AxonMIP and AxonStack.On CellMorph, Meijering had slightly higher Dice and IoU, while MorphoBranch had higher Skeleton F1 and clDice and substantially lower length error.Source: §3.3.1 · pp. 7–8, Table 3 · p. 8

3.3.3. Computational Efficiency

  • Average processing time was 0.189s per AxonMIP image, 0.193s per AxonStack image, and 0.138s per CellMorph image.The complete workflow ran in under one second per image on the evaluated hardware.Source: Table 4 · p. 8, §3.3.3 · p. 9

3.3.2. Ablation Analysis of Structure Extraction

  • On AxonStack, removing multiscale ridge enhancement caused the largest degradation across Dice, Skeleton F1, clDice, and length accuracy; hysteresis also consistently supported structural fidelity.Background correction added gains, while denoising had a modest effect and its removal slightly lowered mean length error.Source: Table 5 · p. 8, §3.3.2 · p. 8

3.3.4. Representative Visualization

  • Two CellMorph examples show enhanced fine processes, preserved major and peripheral branches, and graph-derived branch hierarchies, overlays, and Sholl profiles.The examples represent cells with distinct branching morphologies.Source: Figure 6 · p. 9, Figure 7 · p. 9, §3.3.4 · p. 9

3.4.1. Quantitative Performance

  • Export, metric explanation, safety handling, and analysis execution each exceeded 90% task-level success, while parameter refinement was slightly lower.Capability-level scores were consistently high for intent understanding, routing, state and parameter control, and artifact reporting; execution and safe handling were slightly lower.Source: Figure 8 · p. 9, §3.4.1 · p. 10

3.4.2. Representative Interaction Cases

  • Interaction examples show registered actions for executable requests and capability-boundary handling, including declining 3D reconstruction from a single 2D image without invoking unregistered tools.Image processing and morphometric measurements remained with deterministic modules.Source: Figure 9 · p. 10, §3.4.2 · p. 10

4. Discussion

MorphoBranch integrates fine-structure extraction, skeleton-graph analysis, morphometric measurement, quality-control visualization, and natural-language refinement in a human-reviewable workflow. Across the evaluated datasets, it improved centerline preservation and length-estimation accuracy, with limitations in image dimensionality, topology validation, and user-study scope.

4. Discussion

  • MorphoBranch unifies fine-structure extraction, skeleton-graph analysis, morphometric measurement, quality-control visualization, and LLM-assisted refinement in a human-reviewable workflow.It emphasizes explicit workflow state and provenance alongside natural-language refinement.Source: §4 · pp. 10–11
  • Across AxonMIP, AxonStack, and CellMorph, MorphoBranch improved centerline preservation and length-estimation accuracy over evaluated baselines.Meijering filtering achieved slightly higher Dice and IoU on CellMorph, while MorphoBranch better preserved centerline geometry and length; ablations supported multiscale structural evidence and hysteresis-based extraction.Source: §4 · p. 11
  • The Refinement Engine supports natural-language parameter adjustment, result querying, and workflow control while keeping quantitative analysis deterministic.Adjustments are inspectable and require user confirmation; unsupported requests remain within registered capabilities rather than being delegated to autonomous biological analysis.Source: §4 · p. 11
  • Current scope is limited to two-dimensional images and projections, and CellMorph has few annotated images without exhaustive branch-level topology validation.The LLM evaluation used predefined interaction tasks rather than prospective user studies.Source: §4 · p. 11

5. Conclusions

MorphoBranch combines image analysis, skeleton-graph morphometry, quality-control inspection, and natural-language workflow refinement in a reproducible, human-reviewable process. Evaluations showed improved centerline preservation and length-estimation accuracy over evaluated baselines, and reliable translation of natural-language requests into registered workflow actions.

5. Conclusions

  • MorphoBranch integrates the Morphometry Engine, Workbench Interface, and LLM-assisted Refinement Engine into a reproducible analysis process.The process combines fine-structure-preserving image analysis, skeleton-graph morphometry, quality-control inspection, and natural-language workflow refinement.Source: §5 · p. 11
  • Across AxonMIP, AxonStack, and CellMorph, experiments demonstrated improved centerline preservation and length-estimation accuracy over evaluated baselines.The interaction benchmark showed reliable translation of natural-language requests into registered workflow actions.Source: §5 · p. 11
  • Future work will extend MorphoBranch toward richer structural analysis, broader cellular morphologies, and more comprehensive validation.Source: §5 · p. 11

Ethics statement

The study reports no new animal experiments and describes the provenance and ethics approval associated with the CellMorph images.

Ethics statement

  • No new animal experiments were conducted in this study.Source: Ethics statement · p. 11
  • CellMorph images came from mouse samples associated with the study reported in Wang et al..Source: Ethics statement · p. 11
  • The animal procedures in that study were approved by the University of Electronic Science and Technology of China Ethics Committee for Animal Experimentation.Source: Ethics statement · p. 11
Loading 2610.00860v2…