Skip to content

Synthex for Healthcare AI

Build and Evaluate More Robust Medical Imaging AI

Explore structured workflows for imaging annotation quality, synthetic data research and model robustness evaluation across challenging medical imaging conditions.

Built with these teams in mind

  • Medical imaging AI startups
  • Research institutions
  • Dataset engineering teams
  • ML engineering leads

Research use only. Not a medical device and not intended for clinical decision-making.

Three core problems

Where medical imaging models struggle

Common engineering problems in medical imaging AI, and the approach Synthex proposes to investigate them.

  1. Problem 01

    Medical AI models can fail under unfamiliar conditions.

    Examples

    • Scanner differences
    • Acquisition protocol variation
    • Dataset distribution shift
    • Image noise
    • Difficult lesion appearances

    Proposed Synthex approach

    Controlled robustness evaluation using defined transformations and scenario-specific testing.

  2. Problem 02

    Medical datasets may contain inconsistent annotations.

    Examples

    • Missing regions
    • Incorrect boundaries
    • Reviewer disagreement
    • Suspicious masks

    Proposed Synthex approach

    Annotation quality auditing and human-review prioritisation.

  3. Problem 03

    Training datasets may lack difficult examples.

    Examples

    • Small lesions
    • Rare presentations
    • Difficult imaging conditions
    • Under-represented scenarios

    Proposed Synthex approach

    Researching synthetic augmentation, with careful comparison against appropriate real-data baselines.

Healthcare capabilities

Research capabilities for imaging AI teams

Each capability describes an engineering problem, a proposed workflow and the kind of research output it could produce. All are planned or under research.

  • Platform Vision

    Medical Image Segmentation Evaluation

    Structured evaluation of segmentation models against reference masks.

    Engineering problem
    Aggregate scores hide failures on small or low-contrast structures.
    Proposed workflow
    Per-case overlap and boundary metrics, grouped by lesion size and image condition.
    Example research output
    A per-case evaluation table with flagged failure cases.
  • Platform Vision

    Annotation Quality Auditing

    Find masks that are likely to be wrong before they are used for training.

    Engineering problem
    Label errors and reviewer disagreement silently degrade training and evaluation.
    Proposed workflow
    Agreement checks, geometric sanity checks and model-disagreement signals to rank cases for review.
    Example research output
    An annotation audit report with a prioritised review list.
  • Platform Vision

    Robustness and Distribution-Shift Testing

    Test how models respond to controlled changes in image conditions.

    Engineering problem
    Models trained on one source may behave differently on another.
    Proposed workflow
    Apply defined transformations — noise, contrast, resolution — and compare against the untransformed baseline.
    Example research output
    Performance-change summaries per condition.
  • Research Direction

    Synthetic Medical Imaging Research

    Investigate whether synthetic or augmented examples help where real examples are scarce.

    Engineering problem
    Difficult cases are rare, and collecting more is slow and tightly governed.
    Proposed workflow
    Generate controlled variations and compare models against real-data baselines on held-out real data.
    Example research output
    A comparison study, including cases where synthetic data does not help.
  • Platform Vision

    Dataset Curation

    Assemble research datasets with documented splits, provenance and licence terms.

    Engineering problem
    Leakage between splits and unclear provenance undermine results.
    Proposed workflow
    Patient-level split rules, licence tracking and coverage summaries per split.
    Example research output
    A documented, versioned research dataset manifest.
  • Platform Vision

    Reproducible Model Benchmarking

    Benchmarks that can be re-run exactly as models and data change.

    Engineering problem
    Results are hard to compare when data, code and settings drift.
    Proposed workflow
    Versioned datasets, fixed evaluation configurations and recorded environments.
    Example research output
    A reproducible benchmark with versioned inputs.

How to read status labels

Platform Vision
A planned capability that defines where the platform is heading. Not yet available.
Research Direction
Active exploration of methods and questions. No prototype or results yet; not a product.
Illustrative Demonstration
A visual or interactive explanation using sample content. No live system or real results behind it.
Validated Implementation
Implemented and verified against defined criteria. No capability on this site is presented at this level yet.

Interactive explorer

Medical AI Evaluation Explorer

Choose an imaging condition and an evaluation focus to see how a research evaluation could be structured. Educational illustration — it shows the kinds of metrics researchers might inspect, never measured results.

Illustrative Research Workflow — Not a Clinical Tool

Illustrative Demonstration
Imaging condition
Evaluation focus
  1. Research image
  2. Added Noise
  3. Segmentation Quality
ILLUSTRATION · NOT PATIENT DATA
  • Reference mask
  • Illustrative prediction

Added Noise · Segmentation Quality

Added noise approximates lower-dose or noisier acquisitions. Researchers look for whether small or faint findings stop being detected. Segmentation quality asks how closely predicted regions match the reference, case by case rather than as one average.

Evaluation steps

  1. 01Select held-out research images with reference masks
  2. 02Add synthetic noise at defined, recorded levels
  3. 03Run the segmentation model under a fixed configuration
  4. 04Compare predicted masks with reference masks per case
  5. 05Inspect boundary errors and missed small findings

Metrics researchers might inspect

  • Dice score per case
  • IoU per case
  • Boundary distance (e.g. Hausdorff)
  • Sensitivity by lesion size

Metric names only — no values are shown, and none would establish clinical validity.

Educational illustration for research engineering. It does not process images, produce measurements or support clinical decisions. No patient data is used or collected.

Research workflow

From research dataset to research report

An intended research workflow. Each step produces a recorded artefact, so findings can be traced back to the exact data and settings used.

  1. 01

    Research Dataset

    De-identified research imaging, used under its licence and governance terms.

  2. 02

    Data Preprocessing

    Consistent resampling, intensity normalisation and region-of-interest selection.

  3. 03

    Annotation Quality Assessment

    Audit reference masks and prioritise questionable cases for expert review.

  4. 04

    Synthetic Augmentation ResearchResearch

    Generate controlled variations where difficult examples are scarce.

  5. 05

    Model Training or Evaluation

    Train research models, or evaluate existing ones, on fixed splits.

  6. 06

    Robustness Analysis

    Compare performance across defined conditions and lesion characteristics.

  7. 07

    Research Report

    Document methods, results, limitations and negative findings.

Governance comes first

Research imaging is used only in de-identified form and under its licence terms. This website does not accept uploads or patient data.

Where the engines fit

The Data Engine covers preprocessing, annotation audits and augmentation research; the Evaluation Engine covers training-time evaluation, robustness analysis and reporting.

Technical metrics

Measures used in segmentation research

These metrics describe technical model behaviour on a dataset. On their own they do not demonstrate clinical validity, safety or benefit.

Overlap and detection

Dice Score2 |A ∩ B| / (|A| + |B|)
Overlap between predicted mask A and reference mask B; sensitive to small structures.
Intersection over Union|A ∩ B| / |A ∪ B|
A stricter overlap measure, monotonically related to Dice.
PrecisionTP / (TP + FP)
Of the regions the model marked, how many were correct.
RecallTP / (TP + FN)
Of the reference regions, how many the model found.

Sensitivity and robustness

SensitivityTP / (TP + FN)
The same quantity as recall, the term commonly used in medical imaging — often reported per lesion size.
Scenario-level changem_condition − m_baseline
How a metric changes when an imaging condition changes, reported per case and in aggregate.

Suitability depends on context. TP, FP and FN are true positives, false positives and false negatives; A and B are predicted and reference regions. A metric that suits detection may mislead for segmentation, and class imbalance, label quality and the cost of each error type all change which measures matter.

Research context

Lung CT lesion segmentation and synthetic imaging

A research direction, not a completed study. It explores whether controlled synthetic variation can support lesion segmentation research where annotated examples are scarce.

Research preview

Research Direction
Research question
Can controlled synthetic variation help segmentation models on small or subtle lung lesions, compared with appropriate real-data baselines?
Data context
Public, de-identified research collections such as LIDC-IDRI, used under their data licence terms. Synthex owns no clinical datasets.
Planned approach
Fixed patient-level splits, real-data baselines, synthetic variation as an addition, and evaluation on held-out real images only.
Planned outputs
A methods write-up, reproducible configurations and an honest comparison including negative results.

Pilot positioning

What a focused healthcare research pilot could look like

Example answers to the questions a research or engineering lead would ask. Deliverables are examples of what a scoped pilot could produce — not an operational service.

01What problem does the team face?
Imaging models and datasets whose reliability under varied conditions — and whose annotation quality — is not well understood.
02Why is it difficult to solve?
Real data is scarce and governed, difficult cases are rare, and annotation errors are hard to find at scale.
03How could Synthex address it?
Structured annotation audits, controlled robustness testing and carefully baselined synthetic augmentation research.
04What could a focused pilot produce?
Example deliverableAn annotation audit report, reproducible evaluation results or a research dataset benchmark.
05What evidence would show technical value?
Reproducible per-condition evaluations against agreed baselines, and audit findings confirmed by expert review.

Research collaboration

Evaluate the Next Generation of Medical Imaging AI

We welcome research collaborations on annotation quality, robustness evaluation and synthetic imaging research. Please do not send patient data through this website.

Not a medical device. Nothing on this page is intended for diagnosis or clinical decision-making.