Radiology AI Review System Proposal

A proposal for a sidecar system that receives eligible radiology studies, runs available AI diagnostic pipelines, presents suspicious findings to radiologists, and stores granular review data for audit, analytics, and future workflow optimizations. Official diagnosis and report submission remain in the hospital's existing workflow.

Proposal Intent

This site serves as a memo for technical product stakeholders. We present the system concept: what the radiologist sees, what the backend stores, where asynchronous work happens, what gets audited, and where decisions remain open.

Current framing The AI and ML pipelines will be described and developed separately. This proposal focuses on system integration, modality-limited sidecar activation, radiologist review UX, persistent workflow state, auditability, and the data generated by real usage.

Scope

Image Intake

Receive DICOM studies from hospital imaging systems. Keep raw clinical records immutable from our point of view.

AI Findings

Represent model outputs generically: classifications, confidence scores, bounding boxes, segmentation masks, etc. as well as more high-level summaries.

Human Review

Let radiologists accept, reject, modify, ignore, or separately diagnose findings. We treat disagreement and correction as valuable first-class data we can use for future improvements.

Primary System Principles

Principle	Design implication
Raw medical records are external source-of-truth data	We do not mutate raw DICOM or hospital EHR records.
Radiologist sign-off is the clinical source of truth	AI can suggest, summarize, draft, and prioritize. Final diagnosis and official report submission remain in the radiologist's existing workflow.
AI review activates only where supported	The sidecar appears only for supported modality/model coverage, initially chest CT, mammography, and brain MRI examples.
Radiologist interactions are product data	Corrections to AI findings become analytics inputs for model monitoring, UX evaluation, QA sampling, and future training data curation.
Everything meaningful is evented	Study transitions, AI runs, annotation changes, sidecar draft edits, handoff/status updates, and access events are appended as structured records.

Next recommended section See the System Architecture page for how these principles map to ingestion, queues, storage, AI results, review, reporting, and audit/state tracking.

Sample Dataset Usage

Note: The sample data are PNG/JPG derivatives from Kaggle datasets, not production DICOM studies.

Reader Path

Architecture Modality examples Radiologist workflow Data model Audit and analytics Operations Assumptions