AI Cobb Angle Measurement on a Medical Imaging Platform

Industry

Healthcare, Medical Imaging Software

Geography

California, USA

Cooperation Period

7 Months

Summary

The client operates an AI based medical imaging platform used by radiology groups and orthopedic specialist clinics. The platform reads studies across general radiography, cross sectional imaging, and specialty musculoskeletal workflows. Scoliosis reading depends on Cobb angle measurement, and both the manual measurement burden and the free text report reformatting were slowing the reading workflow.

Discovery sessions ran with reading radiologists, orthopedic surgeons who consume the reports, radiographic technologists who acquire the studies, and the client’s clinical governance group. Five constraints defined the module the platform needed.

Radiologists spent significant time reformatting free text dictation into the structured template each referring specialty required
Manual Cobb angle measurement added several minutes per study, and inter reader variability above 5 degrees on roughly one in six studies triggered clinical callbacks and third read requests
Prior report retrieval relied on manual PACS search, so comparison sections were often written without the previous language the referring surgeon had learned to read
An earlier internal automation attempt held up on adult mild curves and failed on S curves and post surgical hardware, so the reading pool started from low trust
Clinical governance would not approve a module that produced a diagnostic value the radiologist did not sign

Clixlogix engineered a Cobb angle measurement module and a structured reporting stack inside the client’s existing desktop recorder and PACS workflow. Validation ran against a labelled corpus of pediatric and adult standing spine radiographs and EOS studies, stratified by Risser stage, acquisition modality, and curve severity. The Cobb angle module reached a mean absolute error of 1.8 degrees against consensus radiologist reads, with more than 90 percent of measurements within 5 degrees. The same measurement engine was later applied to hallux valgus angle measurement as evidence that the two stage architecture generalizes across musculoskeletal angle tasks, with that module’s validation reported separately. Report turnaround dropped from around 36 hours to under 10 hours and template compliance moved above 90 percent.

The Client

The client is a medical imaging software company serving radiology groups and orthopedic specialist clinics. The platform ingests studies from hospital PACS, presents them in the reader’s preferred viewer, and produces structured reports that flow back to the referring physician through the hospital’s HL7 interface. The scoliosis reading module and the Cobb angle measurement engine described here are one part of a broader musculoskeletal AI portfolio that also includes hallux valgus angle measurement, hip alignment measurement, and long leg length assessment.

Challenge

The platform reached a stage where the reporting workflow itself was the constraint on growth. Adding more radiologists to the reading pool did not reduce turnaround below a floor that sat around 36 hours for scoliosis studies. Five specific problems drove that floor.

Free text dictation to structured template. Radiologists dictated in reading order and then restructured the report themselves or handed it to transcription staff to reformat. The scoliosis template required 14 discrete fields including major curve, minor curve, apex vertebra, Risser sign, skeletal maturity, and comparison to prior. Field completeness across the reading pool averaged around 62 percent, and referring surgeons frequently called back for missing values.
Manual Cobb angle measurement. Radiologists identified the end vertebrae of each curve, placed endplate lines, and computed the angle for both major and minor curves on every study, repeated at every follow up. Inter reader variability crossed the 5 degree surgical planning threshold on roughly 17 percent of studies.
Prior report comparison. The comparison section is high value to the surgeon reading sequentially across visits. Radiologists searched PACS manually and rewrote the comparison text from memory, so the same finding was often described three different ways across three sequential visits.
A prior automation attempt that had failed. The client had previously deployed a landmark detection model that held up on adult mild curves on standing radiographs and failed on S curves with shared vertebrae, on post surgical studies where rods and pedicle screws occluded landmarks, on pediatric studies with smaller less ossified vertebrae, and on EOS full spine studies where the projection geometry differed. Radiologists reverted to manual measurement, so the new module started from low trust with the reading pool.
Regulatory and clinical governance boundary. Clinical governance would not approve a module that produced a diagnostic value the radiologist did not sign. The measurement had to remain decision support, the clinician had to retain responsibility for the value on record, every override had to be captured, and model version had to be traceable to each signed report.

The client’s clinical and product leadership set three success criteria before engineering began. The radiologist workflow inside the existing desktop recorder had to be preserved. Cobb angle measurement had to hit acceptance thresholds across the full clinical distribution including pediatric studies, S curves, severe curves, and post surgical hardware cases. And the module had to sit inside a governance framework the client’s regulatory advisor and clinical lead could sign.

Requirements

The module’s functional scope covered the full measurement path. It had to detect the end vertebrae for the major and minor curves on both standing spine radiographs and EOS studies, propose endplate landmarks on the selected vertebrae, compute the Cobb angles for both curves, and identify the apex vertebra. S curves required double curve detection with a shared transitional vertebra constraint applied where clinically appropriate for the curve type. Post surgical studies required hardware detection and confidence degradation on any occluded landmarks. The radiologist had to be able to override at both the vertebra selection stage and the landmark detection stage. Accepted values had to flow into the structured scoliosis report template as discrete fields and return to PACS as DICOM SR objects with GSPS overlays. Prior patient reports had to be retrieved and used to pre populate the comparison section. Study orders had to arrive over HL7 ORM and signed results had to leave over HL7 ORU.

The non functional constraints tightened the delivery envelope. End to end latency from study open to sidebar populated had to sit under 10 seconds. Every measurement, every override, and every model action had to be captured in an audit log retained per the client’s clinical records policy. PHI had to remain inside the approved deployment boundary, with de identification applied before any data entered the retraining corpus. Every signed report had to trace through its DICOM SR object back to the exact model version that produced the measurement. The module had to run against the DICOM and HL7 profiles the client’s target sites already used, without a custom API contract per site. And validation had to be stratified across patient age, Risser stage, acquisition modality, curve severity, and post surgical status before any production release.

Solution

Clixlogix engineered the Cobb angle module and the structured reporting stack as services that plug into the client’s existing desktop recorder and PACS. The radiologist opens a study, dictates, and signs off the same way they always have. Everything the AI contributes appears inside that flow as pre populated template fields, retrieval assisted comparison text, and a Cobb angle sidebar next to the PACS viewer with per measurement confidence and an override control.

Vertebra selection before landmark detection

The prior automation attempt had gone straight to landmark detection, which is why it broke on S curves. Clixlogix separated the pipeline into two clinically meaningful stages that mirror how a radiologist actually reads a scoliosis film.

Fig 1 – Cobb angle sidebar architecture inside the PACS reading workflow

Stage 1, end vertebra selection. A model trained on labelled studies with radiologist consensus tilt boundaries identifies the upper and lower end vertebra of each curve, applying the clinical rule that the end vertebra is the one whose endplate tilts most toward the concavity. On S curves the stage runs twice, once for the proximal thoracic curve and once for the main thoracolumbar or lumbar curve. Where clinically appropriate for the curve type, the model enforces a shared transitional vertebra constraint that binds adjacent curves through their shared end vertebra. This structural rule alone eliminated a class of measurement errors the prior model had produced.
Stage 2, landmark detection. A landmark model places endplate points on the selected end vertebrae and computes the Cobb angle from the intersection of the two endplate lines. The module reports confidence at both stages, so the radiologist sees whether it is confident about the selected vertebrae, the placed landmarks, both, or neither.

The radiologist can override at either stage. Overriding the vertebra selection triggers a re run of the landmark detection on the new selection. Overriding a single landmark point recomputes the angle without changing the selection. Both override paths write to the training feedback store.

Why this mattered

Separating vertebra selection from landmark detection let the module enforce a shared transitional vertebra constraint on S curves, eliminating a class of measurement errors the prior single stage model had produced and surfacing confidence at the stage where confidence was actually in question.

Fig 2 – The two stage measurement pipeline, vertebra selection before landmark detection

Handling S curves, post surgical hardware, pediatric anatomy, and EOS geometry

Four hard cases required specific engineering.

For S curves, the two stage architecture already reduced the error mode, and the training set was rebalanced to over represent double major and triple curve studies until the model no longer defaulted to a single curve on ambiguous cases.

For post surgical studies with rods and pedicle screws, hardware occluded endplate visibility. Clixlogix added a hardware detection head that flags the study as post surgical, degrades reported confidence, and surfaces a warning in the sidebar that manual placement is recommended. The module still proposes landmarks but never asserts a high confidence measurement on a hardware occluded endplate. Hardware cases remained eligible for assistive landmark proposals but were excluded from the auto accept workflow, and the radiologist manually confirms every measurement on a hardware occluded study before sign off regardless of the module’s reported confidence.

For pediatric studies, vertebral body appearance differs by skeletal maturity, and the Risser sign informs both the report and the model’s expected anatomy. The training set was stratified across Risser stages and re balanced toward younger patients whose vertebral bodies show less ossification. A separate pediatric evaluation slice was reported at every model release.

For EOS full spine studies, the biplanar low dose acquisition geometry differs from a standard AP standing radiograph. Clixlogix added an acquisition metadata parser that reads the DICOM header, identifies the study as EOS, and routes to a fine tuned model head trained on EOS studies specifically. When routing confidence sits below threshold, the parser defaults to the general standing model and flags the acquisition type for radiologist confirmation before the measurement writes to the report. Reporting the acquisition modality on every measurement made the routing decision visible to the radiologist on the sidebar.

Acceptance criteria and clinical validation

Clinical governance and Clixlogix agreed the acceptance criteria before the module went into any radiologist’s live workflow. Validation ran against a hold out corpus of 2,400 studies stratified across the distribution the platform actually reads. The reference standard was the consensus read of two board certified radiologists with adjudication by an orthopedic spine surgeon when the two disagreed by more than 3 degrees.

Acceptance criterion	Threshold agreed with clinical governance	Result at production release
Mean absolute error against consensus reference	Under 3 degrees overall	1.8 degrees overall
Measurements within 5 degrees of consensus reference	Above 85 percent overall	91 percent overall
Pediatric slice mean absolute error (Risser 0 to 3)	Under 3.5 degrees	2.4 degrees
Adult slice mean absolute error (Risser 4 to 5 and skeletally mature)	Under 3 degrees	1.6 degrees
Standing radiograph slice mean absolute error	Under 3 degrees	1.7 degrees
EOS slice mean absolute error	Under 3 degrees	2.0 degrees
Mild curve slice (Cobb 10 to 25 degrees)	Under 3 degrees	2.1 degrees
Moderate curve slice (Cobb 25 to 45 degrees)	Under 3 degrees	1.6 degrees
Severe curve slice (Cobb above 45 degrees)	Under 4 degrees	1.9 degrees
S curve identification accuracy (double major detected as double)	Above 90 percent	94 percent
Post surgical hardware detection recall	Above 95 percent	97 percent
Agreement with orthopedic surgeon read (95 percent limits of agreement)	Within plus or minus 6 degrees	Within plus or minus 5.2 degrees

Fig 3 – Acceptance criteria coverage across the clinical distribution slices

A companion hallux valgus angle measurement module built on the same measurement engine went through the same validation protocol under separate documentation, providing evidence that the two stage architecture generalizes across musculoskeletal angle tasks.

Structured reporting stack and prior report retrieval

The dictation stream from the recorder application routes to Deepgram’s medical speech model, extended with spine and musculoskeletal vocabulary drawn from a corpus of prior sign off reports. LangChain orchestrates the extraction pipeline. As the transcript accumulates, GPT 4 parses the running text into the fields the active template requires, guided by a template selector that picks the correct schema from the study type and referring specialty. A validation stage checks that mandatory fields have been addressed and surfaces any missing field in the recorder pane as a soft prompt before sign off.

Every historical report for the patient is embedded and stored in a Chroma vector store scoped to that patient’s record. When a new study opens, the system retrieves the most relevant prior findings and surfaces the full prior report in a side panel plus a pre populated comparison section written in the phrasing the practice has used across the patient’s history. The system displays the source prior report beside every generated comparison and blocks comparison claims not supported by retrieved prior text. Any statement the model generates without direct grounding in a retrieved prior passage is flagged for radiologist confirmation before it can enter the signed report. The radiologist confirms, edits, or replaces the text.

Standards based integration with PACS, RIS, and EMR

Integration was engineered against clinical standards, so the module deploys into any imaging site the platform already serves without a custom API contract per site.

The Cobb angle module receives the study through a DICOM query and retrieve against the site’s PACS. Measurements return to PACS as a DICOM Structured Report (SR) object encoding the measured angles, the identified end vertebrae, the confidence scores, and the model version that produced them. Visual overlays showing the endplate lines and vertebral labels return as a DICOM Grayscale Softcopy Presentation State (GSPS) object bound to the source study, so the referring surgeon opening the study in any DICOM viewer sees the annotated measurement without needing the Clixlogix client.

Order and result exchange with the hospital information system runs over HL7. Incoming study orders arrive as HL7 ORM messages and drive template selection at the reporting stack. Signed reports leave as HL7 ORU messages carrying the structured template fields, including the Cobb angle SR object reference, so the referring surgeon’s EMR receives the measurement as discrete queryable data alongside the narrative report. Report field mapping between the internal template schema and the receiving EMR’s expected segments was configured per site during onboarding.

Fig 4 – Standards based integration across PACS, HIS, and EMR

Role mapping across the clinical workflow

The module sits inside a workflow with distinct roles whose responsibilities were defined explicitly.

The radiographic technologist acquires the standing spine radiograph or EOS study, confirms patient positioning, and pushes the study to PACS. The technologist has no interaction with the measurement module and no responsibility for the reported value.
The radiologist validates the measurement. The module proposes end vertebra selection, endplate landmarks, and Cobb angle values with confidence scores. The radiologist accepts, adjusts, or overrides at any stage. The signed value on the report is the radiologist’s value, whether the module proposed it or the radiologist replaced it.
The orthopedic surgeon consumes the structured report and the DICOM SR object through the EMR. The surgeon reads the trend across visits to guide bracing and surgical timing decisions. The surgeon does not interact with the measurement module directly.
The admin and QA reviewer monitors the override log. Studies where the radiologist override changed the measured angle by more than 5 degrees are queued for QA review, which feeds retraining priorities.
The clinical governance group reviews measurement drift on a rolling schedule. If any slice of the acceptance criteria table degrades below threshold on new production data, the module is flagged for retraining and the model version in production remains locked until the retrained candidate re passes validation.

Safety, PHI, and governance boundary

Governance boundary

The module is decision support, not autonomous diagnosis. Every measured value on a signed report is a clinician approved value, and the module produces no output that reaches an EMR without a radiologist signature.

PHI stays inside the approved deployment boundary during inference. The measurement engine runs inside the client’s cloud tenant with per site logical isolation, and studies are de identified before any data is used for retraining. The retraining corpus is governed by a data use agreement between the client and each contributing site.

Audit logging captures the model version that produced every measurement, the confidence scores at both the vertebra selection stage and the landmark stage, whether the radiologist accepted or overrode at each stage, the final signed value, and the timestamp and user identity for every action. The audit log is retained per the client’s clinical records retention policy and is available for regulatory inspection.

Model version traceability runs end to end. Every DICOM SR object written to PACS carries the model version that produced the measurement. Every signed report references the model version through the SR link. If the module is retrained and a new version deployed, historical reports remain traceable to the version that produced them.

Regulatory positioning treats the Cobb angle module as clinical decision support with clinician in the loop, consistent with the FDA’s guidance on radiological computer assisted detection and measurement software. Comparable commercial modules in this space commonly position around CE marking, Health Canada, FDA cleared detection modules, or peer reviewed clinical validation, so regulatory evidence and traceability sit as a buyer expectation the platform had to meet. The module was not deployed to production until validation against the acceptance criteria table had been signed by the client’s clinical lead and regulatory advisor.

Results

The Cobb angle module and the structured reporting stack are complete and running in production at pilot sites. Scoliosis reads flow through the module and into the referring surgeon’s EMR through DICOM SR and HL7 ORU. Radiologists sign every measurement. Adoption did not require classroom training since the recorder workflow itself did not change. The six results below are measured across the first reading waves after production rollout.

1.8 Degree Mean Absolute Error Against Consensus

The measurement engine reached a mean absolute error of 1.8 degrees against the consensus of two board certified radiologists, down from over 4 degrees on the prior automation attempt and inside the 3 degree threshold clinical governance had agreed as the release condition.

91 Percent of Measurements Within 5 Degrees

Across the validation corpus of 2,400 studies, 91 percent of the module's Cobb angle measurements sat within 5 degrees of the consensus reference read, clearing the 85 percent acceptance threshold agreed with clinical governance.

Inter Reader Variance Cut from 17 Percent to Under 7 Percent

Inter reader Cobb angle variance above the 5 degree surgical planning threshold fell from 17 percent of studies before the module to under 7 percent after rollout, measured on a rolling dual read subset of 400 studies per quarter with post surgical hardware studies excluded.

Report Turnaround Reduced from 36 Hours to Under 10

Average scoliosis report turnaround fell from around 36 hours to under 10 hours after production rollout, from measurement pre population, structured template completion driven by field extraction from dictation, and prior comparison retrieval, each contributing a distinct share.

Template Field Completeness Rose from 62 Percent to Above 90 Percent

The 14 field scoliosis reporting template went from 62 percent field completeness across the reading pool to above 90 percent after the reporting stack went live, with the validation stage flagging missing fields before sign off.

Referring Surgeon Callback Rate Reduced by More Than 70 Percent

Referring surgeon callbacks for missing scoliosis report fields fell by more than 70 percent after the module went live, driven by higher template completeness at sign off and discrete field delivery through DICOM SR into the referring EMR.

Technologies and Tools

Category	Tools and Platforms
Cobb Angle Measurement	End vertebra selection model (keypoint based vertebra selection with tilt scoring); endplate landmark detection model (heatmap based landmark regression over selected vertebrae); hardware detection classifier (rods and pedicle screw detection with confidence degradation); EOS routing parser (DICOM metadata with image based fallback); ONNX Runtime (GPU accelerated inference inside approved deployment boundary)
Model Training	PyTorch (training pipeline with stratified slices for Risser stage, acquisition modality, curve severity, and post surgical status); augmentation pipeline (pediatric and severe curve weighting); validation harness (MAE, within 5 degree rate, slice level acceptance checks)
Structured Reporting	Deepgram medical speech model (spine and musculoskeletal vocabulary tuning); GPT 4 (`JSON` schema constrained extraction and grounded comparison drafting); LangChain (template selection, extraction orchestration, validation prompts); scoliosis report schema (14 structured fields with mandatory completion checks)
Prior Report Retrieval	Chroma (patient scoped prior report vector store); clinical text embeddings (radiology report retrieval); source linked comparison panel (prior report display beside generated comparison); unsupported claim blocking (comparison text constrained to retrieved prior content)
Clinical Integration	`DICOM Query and Retrieve` (study intake from PACS); `DICOM SR` (Cobb angle values, end vertebrae, confidence scores, model version); `GSPS` (editable measurement overlays in supported DICOM viewers); `HL7 ORM` (inbound orders); `HL7 ORU` (signed structured results to EMR); per site field mapping (receiving EMR segment configuration)
Workflow Integration	Desktop recorder integration (audio routing and transcript return); PACS sidebar module (study context from active DICOM identifier); structured report pane API (accepted values inserted into scoliosis template); override controls (accept, adjust, override actions captured for QA and retraining)
Governance and MLOps	Model registry (training snapshot, validation slices, deployment metadata); version traceability (model version written into `DICOM SR` and signed report reference); override feedback store (high magnitude edits prioritized for review); drift monitoring (monthly slice level review by clinical governance)
Security and Compliance	Tenant level isolation (per site logical boundary); PHI residency (inference inside approved deployment boundary); de identification pipeline (retraining corpus prepared under site data use agreements); audit logging (measurement actions, user identity, timestamp, final signed value); regulated software controls (clinician sign off, validation evidence, traceability, regulatory advisor review)

Why This Engagement Worked

Four engineering decisions carried the outcome.

The measurement pipeline separated vertebra selection from landmark detection, so the module’s reasoning followed how a radiologist actually reads a scoliosis film. The prior automation attempt had gone straight to landmarks and had failed on the cases that mattered most. The two stage design let the module handle S curves through a structural rule that eliminated the single curve default error class, and it let confidence scores surface at the stage where confidence was actually in question.

Acceptance criteria were agreed with clinical governance before validation and reported at the slice level the clinical distribution demands. The pediatric slice, the EOS slice, the severe curve slice, and the post surgical hardware slice each had its own threshold and its own reported result. Aggregate accuracy hid the failure modes that mattered clinically, so aggregate accuracy was never the number the module was released on.

Integration ran against clinical standards. DICOM SR carried the measurement, GSPS carried the overlay, HL7 ORM and HL7 ORU carried orders and results, and the referring surgeon received the measurement as discrete EMR data. Any imaging site the platform already served could receive the module without a proprietary API contract.

Governance was engineered in from the start. The module is decision support, the clinician signs the value, every override is captured, every measurement carries a traceable model version, and the clinical governance group reviews drift on a rolling schedule. Regulatory positioning followed FDA guidance for clinical decision support software from the first design review, not as a retrofit before launch.

Services Delivered

Digital Engineering, AI Software & Application Development, Computer Vision Engineering, Speech Recognition & Structured Reporting, DICOM & HL7 Integration, MLOps & Clinical Validation, Consulting Service

Team Composition

AI Product Architect, Computer Vision Engineers, ML Engineers, Medical Informatics Specialist, Backend & Integration Engineers, Clinical QA Analyst, Regulatory Advisor

Have a question for our team or need help with your project?

Our team can share client references, scope your project, and answer any question about your delivery.

Related Case Studies

More engagements where our delivery teams shipped similar outcomes for clients across industries. Read on for context on the patterns we reused, the trade offs we navigated, and the metrics that landed in production.