Start a project
AI intelligence system

ClinicalInsight

From image finding to evidence-linked insight.

A clinician-facing decision support system that combines medical imaging, patient history and retrieved literature into findings a radiologist reviews, with the model positioned as a second reader rather than a decision maker.

Healthcare Clinical Imaging Diagnostics Support Multimodal Human in the loop
Sector
Healthcare & Clinical Imaging
Year
2025
Duration
12 months
Team
9 people
Platforms
Web · PACS integration
Status
In clinical evaluation
ClinicalInsight — visual identity for the case study
Second reader System role Never the decision maker
Longitudinal Comparison scope Against the patient's own priors
Cited Literature support Retrieved, never recalled
Full Audit trail Every acceptance and rejection

01 — The problem

Imaging AI tends to fail in the same way. A model outputs a probability for a finding, the clinician has no way to interrogate it, and it becomes another number to either trust blindly or ignore entirely. Most are ignored.

The clinically useful work is not detection alone. It is relating a current finding to the patient's prior imaging, their history, and what the literature says about that combination — which is exactly the part that gets left to the clinician's memory at the end of a long list.

02 — What we did

We designed the system as a second reader that shows its working. Every finding comes with the region it rests on, the comparison against prior studies, the relevant history, and the literature that informs interpretation, all inspectable before the clinician forms a view.

The workflow is explicitly human-in-the-loop, and structurally so. The system cannot write to the report. It prepares, evidences and prioritises; a clinician reviews, accepts or rejects, and every one of those decisions is recorded with its rationale.

A probability without a reason is another number to ignore at the end of a long list.
Interface

What it looks like in use

Two of the screens that carry the most weight in daily use, rebuilt here from the production design system.

clinicalinsight.aurezalabs.com/study/44182
Study 44182 · CT chestSECOND READER
Nodule · RUL · 8.2mmREGION SHOWN
Prior 2023-11 · 6.1mm+34% INTERVAL
Literature · 4 passagesCITED
Clinician decision requiredPENDING
System cannot write to report

A second reader that shows its workingEvery finding opens onto the region it rests on, the prior comparison and the literature that informs interpretation.

clinicalinsight.aurezalabs.com/monitoring
Model performance · liveMONITORED
Findings accepted81%
Findings rejected19%
Drift alerts · 90 days0
Weekly agreement

Drift watched in serviceModel behaviour is compared continuously against clinician decisions, so degradation is detected rather than assumed away.

Capabilities

What the system does

Grouped by the job each set of capabilities exists to do, rather than by which team built it.

Imaging

01 · Reading the study

Medical image ingestion

DICOM ingestion from PACS with modality-aware preprocessing and study-level organisation.

Anatomical segmentation

Structures segmented so findings are described in anatomical terms rather than pixel coordinates.

Abnormality detection

Candidate findings localised with the supporting region highlighted for direct inspection.

Calibrated confidence

Confidence expressed against measured operating characteristics rather than a raw model score.

Clinical context

02 · Where the value is

Longitudinal comparison

Current study compared against the patient's own prior imaging with change quantified and shown.

History integration

Relevant history, medications and prior results surfaced alongside the finding they bear on.

Literature retrieval

Passages retrieved from trusted sources for the specific finding and patient context, always cited.

Differential support

Considerations presented as a structured set with the evidence for each, not a ranked verdict.

Workflow and governance

03 · Safety by construction

Clinician review workflow

Every finding requires explicit acceptance or rejection. Nothing propagates unreviewed.

Worklist prioritisation

Studies ordered by finding urgency so time-critical cases reach a reader sooner.

Complete audit trail

Every model output, clinician decision and rationale recorded immutably.

Performance monitoring

Live tracking of model behaviour against clinician decisions to detect drift in service.

Architecture

How it is put together

Layer by layer, with the reason each one exists — because the reason is usually the interesting part.

01PACS integration

Standards-based DICOM integration with the imaging estate, operating alongside existing workflow rather than replacing it.

  • DICOM
  • HL7
  • DICOMweb
02Imaging models

Segmentation and detection per modality and body region with calibrated confidence outputs.

  • PyTorch
  • MONAI
  • Calibration
03Longitudinal engine

Registration of current against prior studies with quantified interval change.

  • Image registration
  • Change quantification
04Knowledge retrieval

Retrieval over curated clinical literature scoped to finding and patient context, with citation enforced.

  • pgvector
  • Curated corpus
  • Citation gate
05Review application

Reading workspace with region overlay, prior comparison, evidence panel and the decision audit trail.

  • React
  • Cornerstone
  • Audit store

Technology

AI & Imaging

  • PyTorch
  • MONAI
  • Registration
  • pgvector

Integration

  • DICOM
  • DICOMweb
  • HL7 FHIR

Backend

  • Python
  • FastAPI
  • PostgreSQL
  • Celery

Frontend

  • React
  • TypeScript
  • Cornerstone.js
Outcome

What changed

Measured against how the operation ran before, not against a benchmark chosen after the fact.

  • The model is inspectable. Every finding opens onto the region, the prior comparison and the literature behind it.

  • Prior studies are always in view. Interval change is computed rather than left to recall.

  • The clinician remains the decision maker. The system cannot write to the report, by construction.

  • Drift is monitored in service. Model behaviour is continuously compared against clinician decisions.

How it ran

  1. Months 1-3

    Clinical and regulatory framing

    Defined the intended use, the second-reader boundary and the evaluation protocol with the clinical team.

  2. Months 4-6

    Integration

    PACS and record integration built to standards so the system sits inside existing workflow.

  3. Months 7-9

    Models and longitudinal engine

    Segmentation, detection, calibration and prior-study registration.

  4. Months 10-11

    Evidence and review

    Literature retrieval with citation gating and the clinician review workspace.

  5. Month 12

    Clinical evaluation

    Shadow-mode operation with performance monitored against clinician decisions.

Related work

From the same practice

All fifteen projects

Manufacturing & Production Quality

LineTrace

From defect detection to defect causality.

A multimodal system that correlates line video, machine telemetry, batch records and operator actions to explain why a defect happened, not merely that it did..

2025 · 10 months
Telecommunications & Network Operations

NetCause

Not alarm detection. Failure hypothesis reduction.

A correlation engine that turns an alarm storm into a short ranked list of probable root causes by reasoning over network topology and fault propagation rather than alarm counts..

2025 · 8 months
Oil & Gas · Asset Integrity

PipeIntegrity

From a single inspection to longitudinal asset integrity.

A vision system that detects and localises pipeline deterioration, then tracks each defect across inspection cycles so maintenance is prioritised by progression rather than by appearance..

2024 · 9 months
Something similar?

Running into the same problem ClinicalInsight solved?

Tell us what you run. We will reply within two business days with what we would build for your situation — and, just as usefully, what we would leave out.

Start a conversation More work