Zhongjie Huang
中文

MSc PROJECT / 2026 · UNIVERSITY OF SHEFFIELD

One boundary.
Many consequences.

Investigating how cell boundaries shape measurements, and what those measurements can support.

The 30-second overview Explore the details
4Existing pipelines
3Levels of evidence
H&EHistology images
THE MEASUREMENT CHAIN
Original synthetic cell regions. Illustrative geometry only, not patient tissue or model output.
Original illustration · not model output02 / 02

01 / THE OVERVIEW

The project, at a glance.

30-second read
01

Why

A tiny change in a cell outline can change its measured size and shape.

02

Question

Does better segmentation also make measurements and predictions more useful?

03

Approach

Compare existing segmentation methods, inspect morphology and evaluate downstream features.

04

Evidence

Check three things separately: reference agreement, measurement support and predictive value.

05

Contribution

A comparative research workflow with traceable processing and explicit evaluation boundaries.

06

Lesson

A convincing outline is not enough. Each downstream claim needs its own evidence.

02 / DIFFERENT BOUNDARIES

Same task. Different assumptions.

Four alternatives, not four stages in one model.

Original synthetic cell regions. Illustrative geometry only, not patient tissue or model output.SYNTHETIC GEOMETRY · NOT MODEL OUTPUT

Nucleus → expansion

QuPath

Detect nuclei, then grow a surrounding territory. A common expansion distance makes the assumption easy to inspect.

Assumption to check: A fixed expansion cannot capture every irregular outer boundary.

03 / THREE LEVELS OF EVIDENCE

A good outline is not the finish line.

01 / SEGMENT

Does it match?

Agreement with an expert reference.

IoU · PQ · Boundary F1
≠
02 / MEASURE

Is it a reliable measurement?

Area error and the objects retained.

Area · Shape · Eligibility
≠
03 / PREDICT

Does it add useful information?

Incremental value over a fair baseline.

Image vs morphology vs fusion

This is a methods-focused public overview. The MSc assessment and permissions are not confirmed in the supplied records, so numerical findings, original figures and project-specific configurations are withheld — not hidden in the page source.

04 / TECHNICAL DEEP DIVE

The full workflow, without the wall of text.

01Research questions & ownershipA comparison of existing methods, not a new segmentation architecture.

My contribution was to compare and connect established methods: turn images into labelled objects, objects into measurements, and measurements into features that can be evaluated. The original segmentation architectures belong to their respective authors.

Recover objects

Are individual reference cells found, and do their outlines overlap?

Measure objects

Are the retained objects measured consistently, and which objects were excluded?

Use the features

Does morphology add information beyond image features under a fair comparison?

02Data, calibration & separationDifferent data answer different questions.

Reference-based evaluation

CytoNuke provides public expert-reviewed nuclear and whole-cell annotations. These are different targets. A nuclear mask is not automatically a whole-cell mask.

Downstream tissue analysis

Oral histology supports analysis of inferred morphology and grade-related features. Without reference cell outlines, method agreement alone cannot establish anatomical accuracy.

Keep related images together when splitting data. Verify physical scale before measuring area. Preserve instance labels when resizing masks. Record provenance, grouping assumptions and exclusions rather than inferring them from a reorganised file path.

Cohort-specific counts, case identities, image files and data manifests are intentionally absent from this public overview.

03The four method familiesInputs, mechanisms, assumptions and configuration choices.
QuPath

Nuclear detection and geometric expansion

Stain separation and watershed locate nuclei. An expansion distance estimates a surrounding cell territory.

Configuration questions: Detection threshold, physical expansion radius and nuclear size limits.

Cellpose-DINO

Features, foreground and flow

A DINO representation supports prediction of foreground and pixel flows; flow-based recovery separates instances.

Configuration questions: Fine-tuning scope, assumed diameter, foreground threshold, flow filtering and minimum size.

Cyto R-CNN

Nucleus-conditioned whole-cell masks

Multi-scale features and nuclear proposals supply regions for whole-cell mask heads.

Configuration questions: Trainable cell heads, proposal scores, non-maximum suppression and retained mask area.

CSGO

Nuclei, membrane estimate and watershed

Nuclear seeds and a predicted membrane landscape guide the allocation of cell-like territories.

Configuration questions: Nuclear detection, membrane threshold, territory size and foreground constraints.

A comparison is between configured pipelines: pretraining, adaptation, parameter searches and post-processing all matter. It cannot isolate architecture alone. Project-specific settings and training records are not distributed in this public build.

04The complete research workflowFrom data checks to evidence, one stage at a time.
  1. 01

    Prepare

    Audit images, units, annotations and related-sample groups.

  2. 02

    Adapt

    Reuse pretrained methods; state which components are updated.

  3. 03

    Select

    Use development data for settings and checkpoint selection.

  4. 04

    Validate

    Check mask shape, labels and selection evidence.

  5. 05

    Test

    Keep selected settings fixed for the final comparison.

  6. 06

    Transfer

    Apply each alternative independently to the target images.

  7. 07

    Measure

    Compute shape, size, eligibility and case summaries.

  8. 08

    Represent

    Extract frozen image embeddings as a separate feature path.

  9. 09

    Compare

    Evaluate image, morphology and fusion branches fairly.

  10. 10

    Interpret

    Audit outputs and explain what the evidence supports.

Shared output contract: a two-dimensional integer mask, with zero as background and a positive label for each object. This makes downstream analysis comparable without pretending the methods produce the same kind of anatomical boundary.

05Segmentation & measurement evaluationWhy one good score cannot validate the whole pipeline.
Overlap
IoU = |P ∩ R| / |P ∪ R|

P is a predicted object and R a reference. One-to-one matching prevents multiple fragments from claiming the same reference cell.

Detection and overlap together
PQ = Σ IoU / (TP + ½ FP + ½ FN)

Sum over accepted matched pairs. TP, FP and FN count matched predictions, unmatched predictions and missed references. Report the matching policy alongside the score.

Conditional area error
e = |Apred − Aref| / Aref

This compares areas only for eligible matched objects. Missed cells do not contribute. A low error can therefore describe a selected subset rather than the full tissue population.

Count retained objects as well as reporting their errors. Aggregate at the independent sampling level; thousands of cells from a few patients are not thousands of independent patients. Boundary tolerance, border rules and averaging must remain explicit.

06Morphology: an interactive explanationChange a synthetic outline and inspect the consequences.

Illustrative geometry. Dashed: smooth reference. Assumed scale: 0.5 μm per simulated pixel.

Move the boundary. Watch the measurements.

Area
733.8μm²
Perimeter
100.9μm
Circularity
0.9070–1

Computed from this polygon only. No patient data or dissertation measurements are used.

Physical area
A = Npixels × s2

s is physical length per pixel; area depends on its square.

Circularity
C = 4πA / P2

A is area and P perimeter. In continuous geometry, a circle has circularity 1. Pixel-based perimeter estimates may differ.

Size, perimeter, circularity, eccentricity, solidity, aspect ratio and neighbour distance describe different aspects of a retained object. Case-level summaries must also record counts, border exclusions, density and heterogeneity. A change in summary can reflect a different object population, not only a different boundary.

07Downstream classification & validationImage features, morphology and fusion under comparable splits.

Image-only

Frozen Phikon-v2 embeddings provide a reference feature path from the tissue image.

Morphology-only

Each segmentation method supplies its own object-derived feature block.

Fusion

Combine the two feature types, then compare against the same image-only baseline.

One image branch, four morphology branches and four fusion branches form nine comparisons. Regularised logistic regression provides a controlled classifier rather than another large neural network.

  1. 01Outer groupsHold related cases together
  2. 02Inner selectionTune within training groups only
  3. 03Fold-local transformsFit scaling and PCA on training data
  4. 04Held-out predictionsCompare the same cases and folds
Balanced accuracy for a binary task
BA = (Sensitivity + Specificity) / 2

This gives equal weight to the recall of each class. AUROC and macro-F1 provide complementary views, not substitutes for independent data.

08Uncertainty, source effects & limitsAsk what else might explain an apparently good result.

Paired comparisons

Compare fusion and image-only predictions on the same held-out cases. Ask whether an observed difference is practically meaningful.

Conditional uncertainty

Resampling fixed out-of-fold predictions describes uncertainty conditional on that fitted workflow. It does not include all variation from retraining.

Source controls

A source-proxy baseline can expose a shortcut. A strong result is a warning to investigate, not proof of a specific scanner or staining cause.

Pretraining provenance

Project-level separation cannot prove independence from unknown upstream pretraining exposure. Document what can and cannot be checked.

05 / CRITICAL REFLECTION

From a convincing output to a defensible claim.

This project changed how I think about evaluation. A plausible boundary or a strong score is only one piece of evidence. Measurements, data sources and downstream decisions each need their own checks.
THEN

Compare segmentation scores

WHAT I NOW SEE

Detection, area fidelity and predictive utility differ.

NEXT TIME

Validate each level with its own endpoint.

THEN

Measure inferred contours

WHAT I NOW SEE

Eligibility changes which objects the result describes.

NEXT TIME

Report border rules, missing objects and measurement support.

THEN

Use grouped evaluation

WHAT I NOW SEE

Grouping cannot settle unknown pretraining exposure.

NEXT TIME

Audit provenance and seek external, independent cohorts.

THEN

Compare feature branches

WHAT I NOW SEE

Source composition can be mistaken for biological signal.

NEXT TIME

Test source controls and validate across acquisition settings.

THEN

Report uncertainty

WHAT I NOW SEE

A fixed-prediction interval omits retraining variation.

NEXT TIME

Repeat the whole workflow and add independent references.

These are current reflections and proposed improvements, not claims that the additional experiments have already been completed.

↗Sources, credits & publication boundariesWhat was reviewed, what is public, and why some materials stay private.

The dissertation, source archive and stored aggregate outputs were reviewed to avoid inaccurate project claims. Only a conceptual methods overview is published here. No private records or unpublished numerical findings are embedded.

Policy review: 4 November 2026. Assessed student work may be shareable, subject to ownership, collaboration, module, third-party and data conditions. The general policy is not project-specific permission. This site distributes no dissertation PDF, original research source, model weights or patient images.

Diagrams are original, illustrative geometry. They do not reconstruct patient images or show new model experiments.

Project directory

All projects