Aaron 乐
Work / PROJECT 04

RadIQ: From Radiology Reports to Quality Management and Clinical Collaboration

Voice-generated reports, multimodal quality checks, and diagnostic risk alerts. Deployed and in use at three hospitals, with positive results.

Iterating Healthcare AI Medical Imaging Report Quality Model Engineering

One report connects more than radiology

Radiology is the starting point. Quality management, clinical collaboration, and patient care complete the picture.

  • Radiologists: less repetitive writing; earlier identification of missing information.
  • Department heads: clearer quality issues and improvement priorities.
  • Quality and medical affairs teams: track findings, reviews, and corrective action.
  • Treating clinicians: clear reports and timely handoffs of important risks.
  • Patients: less waiting, timely attention to significant findings, and clearer next steps.
  • Hospital IT: straightforward integration, controlled data access, and manageable models.

Highlights: multimodal AI, deployed at three hospitals

The multimodal AI reporting tool is deployed and in use at three hospitals, with positive results.

Core capabilities include:

  • AI voice-to-report generation: turn dictated findings into structured drafts.
  • Multimodal report quality checks: flag omissions, inconsistencies, and items requiring review.
  • AI diagnostic risk assessment: highlight risks for physician review.
  • Quality management and clinical collaboration: trace quality findings, notifications, and actions.

The workflow: less repetition, traceable review

  • Reporting: voice, text, and images → a draft report.
  • Quality: daily report synchronization → full review → human confirmation.
  • Collaboration: risk alert → clinical notification → recorded acknowledgment.
  • Responsibility: AI drafts and flags; physicians decide and sign.
Report drafting and quality checks in one view; key measurements await confirmation.
Report drafting and quality checks in one view; key measurements await confirmation.

Reporting: from dictation to signature

  • Flexible input: combine dictation, typing, and uploaded images.
  • Structured drafts: organize findings by anatomy; flag missing elements and RADS classifications.
Voice, text, and images feed into the findings section.
Voice, text, and images feed into the findings section.
  • Iterative editing: request changes in plain language; save, compare, and restore versions.
Previous drafts can be reviewed, checked, and restored.
Previous drafts can be reviewed, checked, and restored.
  • Discuss before drafting: summarize a case discussion into findings with one action.
The medical assistant can transfer a discussion summary into the report input.
The medical assistant can transfer a discussion summary into the report input.
  • Accepting is not signing: send the draft to the hospital system for physician review and signature.
Accepted drafts enter the review workflow.
Accepted drafts enter the review workflow.
  • Reading-room support: switch between light and dark modes.
Dark mode for the reading room.
Dark mode for the reading room.

Quality management: beyond spot checks

  • Full review: check all reports included in the daily sync; actual coverage depends on synchronization and task execution.
  • Visible reasoning: show matched rules, deduction reasons, and suggested changes.
  • Human review: AI flags issues; physicians confirm scoring decisions.
  • Continuous improvement: aggregate recurring issues to inform departmental and hospital quality management.
Daily review: pending reports and quality distribution. Demonstration data.
Daily review: pending reports and quality distribution. Demonstration data.

Eight quality dimensions:

  • Report structure and identifier accuracy.
  • Findings–conclusion consistency and inclusion of positive findings.
  • Critical findings and RADS compliance.
  • Measurement plausibility and follow-up recommendations.
Compare the original report with quality findings.
Compare the original report with quality findings.
  • Configurable rules: adjust scoring, activation, and evaluation methods; retain change records.
Centralized quality-rule management.
Centralized quality-rule management.
  • Trackable insights: review dimension-level performance and improvement suggestions.
Department overview and AI quality insights. Demonstration data.
Department overview and AI quality insights. Demonstration data.

Clinical collaboration: alerts need follow-through

  • Prominent critical-finding banners do not dismiss themselves.
  • Connect clinical notification, acknowledgment, and action records.
  • System records do not replace phone calls or hospital response procedures.
Critical-finding alert and clinical notification entry point.
Critical-finding alert and clinical notification entry point.

Access and responsibility

  • Physicians: write reports and see their own quality results.
  • Department heads: see department summaries and quality distribution.
  • Administrators: manage accounts, permissions, and models—not clinical report details.
  • Quality, medical affairs, and collaborating teams: access depends on hospital authorization and system integration.
Physicians see only their own quality results.
Physicians see only their own quality results.
Role-based permissions with traceable changes.
Role-based permissions with traceable changes.

Hospital IT: manageable models, reusable capabilities

  • Replaceable models: configure report generation, speech recognition, quality checks, and insights separately.
  • Controlled upgrades: candidate switching, staged rollout, and connectivity tests.
Centralized model and parameter configuration.
Centralized model and parameter configuration.
  • Evaluate before choosing: blind comparisons for report models; side-by-side speech benchmarks.
Blind evaluation of report-generation models. Demonstration data.
Blind evaluation of report-generation models. Demonstration data.
Speech-recognition latency and error-rate comparisons. Demonstration data.
Speech-recognition latency and error-rate comparisons. Demonstration data.
  • Maintainable terminology: manage domain vocabulary and weights by group.
Medical terminology vocabulary management.
Medical terminology vocabulary management.
  • Reusable evaluations: use the same cases for human and model-based assessment.
Evaluation dataset management. Simulated cases.
Evaluation dataset management. Simulated cases.
  • Reusable speech services: expose text-to-speech, audio recognition, and real-time transcription to other hospital systems.
Hospital speech-service API demonstration.
Hospital speech-service API demonstration.

Deployment and boundaries

  • On-premises: keeping data in the hospital requires local models and appropriate network controls.
  • Physician responsibility: AI provides drafts and risk prompts, not final diagnoses.
  • Authorized access: a management need does not imply cross-department access is enabled.
  • Patient value: waiting times and care coordination are goals to evaluate—not proven outcomes or an existing patient portal.
  • Demonstration data: screenshots are redacted; sample statistics are not hospital operating results. The product UI shown is in Chinese.
  • Intellectual property: independently developed; V1.0 software copyright registration held by 北大医疗管理有限责任公司 (Peking University Healthcare Management Co., Ltd.).

Start with one report. Connect quality management and clinical collaboration. Keep patient care as the goal.

Screenshot

Choose Original size, then scroll in any direction to see details.