MedSight · AI triage for radiology, with audit-chain provenance
Get critical scans to the top of the worklist faster, with a tamper-evident record for every AI call. An assistive triage and CADt platform for hospital radiology departments, hospital chains, diagnostic networks and public-health screening. Six P0 study types — chest X-ray, brain MRI, mammography, OCT, skin lesion and Parkinson's spiral — each flagged with a Likert confidence, a Grad-CAM overlay and an Ed25519 audit-chain entry engineered to produce a tamper-evident, BSA §63-aligned evidence bundle (hash, algorithm, chain of custody, device identity, operator). Every high-stakes finding routes through two-radiologist concurrence before it reaches the EMR. Assistive, not autonomous — the radiologist owns the diagnosis. Built under a triple-track CDSCO + US FDA 510(k) + EU MDR regulatory file (regulatory track, in progress).
Maturity: reference implementation / pilot-ready. Production models and live PACS connectors land at the CDSCO chest-X-ray submission milestone; the dev/CI suite exercises deterministic stub adapters today. Available in early access — request access below.
01 — Who it's for
At incumbent chain-licence pricing, full multi-indication coverage across a large hospital group can consume most of the AI-radiology budget — so the flagship sites get AI and the rest get nothing. MedSight is built for an India-domestic cost structure, with a multi-indication bundle and a single product with preset-toggled compliance, so a chain can cover every hospital, not just the showcase ones.
— ICP · 01 · Radiologist
Senior consultant in an 11-radiologist department, 525 scans/day across 12 modality units, ~38-minute median time-to-read for a critical CXR at 2 AM. Wants critical findings surfaced to the top of the worklist, a Grad-CAM heatmap to show the referring physician, and an audit entry that can be cited if a case escalates to NMC or Consumer Forum.
— ICP · 02 · Hospital chain CIO
Group CIO across ~70 hospitals on a fixed AI-radiology budget — a handful of sites on a competitor, the rest on no AI radiology at all. Wants chain-wide rollout inside budget, DPDP + NABH + JCI compliance documentation, and one vendor relationship instead of seven.
— ICP · 03 · Diagnostic network
National diagnostic chain, ~35-radiologist network, hundreds of collection centres, thin per-scan margins. Procures per-scan SaaS, not capex. AI triage has to stay within those thin per-scan margins — MedSight is built to operate within that envelope.
— ICP · 04 · Public health / insurer
Senior underwriting manager at a life insurer, ~4,500 pre-policy CXRs/month read manually today. Wants AI triage at a fraction of the manual cost — with explicit consent, model-version logging, 7-year retention, and a Grad-CAM explanation available to the applicant on request — per IRDAI Guidelines 2024 §11.2.
02 — How it works
The MedSight edge appliance sits on the hospital LAN. It pulls or receives DICOM, de-identifies at the edge boundary before any cloud uplink, runs the per-indication model, raises a triage flag and a worklist priority, hands the prediction to the Ed25519 audit chain, and exports a FHIR R4 DiagnosticReport back to the PACS / RIS / EMR — with a Provenance resource anchored to the chain hash. Nothing leaves the site in identified form.
For any PACS or modality
DICOM Q/R SCU (C-FIND / C-MOVE) pulls studies, or a STORE SCP listener receives instances pushed by the modality. The same header representation flows straight into de-identification.
For DPDP + HIPAA data-residency
DICOM PS3.15 Basic Confidentiality Profile + Clean Pixel Data + Retain Patient Characteristics for fairness analysis. All 18 HIPAA Safe Harbor identifiers removed and verified; residual-PHI failure blocks the uplink.
For per-indication triage
A PaliGemma 2 fine-tune (deterministic stub in dev/CI) produces findings, each with a 1-5 Likert confidence, a continuous probability, an optional bounding box and a Grad-CAM overlay reference. Target: CXR clears in ≤ 5 s p99 on a Jetson AGX Orin (design goal, not yet independently measured).
For 2 AM on-call worklists
A finding at the Likert threshold or any CRITICAL-severity finding elevates worklist priority and fires the critical-finding alert (SMS first, IVR escalation). Targets: priority elevation within ≤ 10 s, alert within ≤ 120 s p99 (design goals, not yet independently measured). Every elevation and reversal is logged with a closed reason code.
For NMC §3.6 high-stakes findings
Likert ≥ 4 or CRITICAL severity mandates a second, distinct radiologist sign-off before the result reaches the EMR. Discordance escalates to a senior tie-breaker; the second operator can never equal the first (two-operator rule, enforced in the type system).
For medico-legal evidence
Every prediction, concurrence and access appends an Ed25519-signed, hash-linked entry to the per-tenant audit chain. Results export as FHIR R4 DiagnosticReport + Observation + Provenance; the BSA §63-aligned evidence bundle is one CLI call away. Admissibility in any proceeding is determined by the court.
03 — Supported study types
Each indication is a standalone triage pipeline with its own closed clinical vocabulary, its own severity policy and its own per-indication clinical-validation gate. The triage-influencing indications are IEC 62304 Class C software; the advisory-only Parkinson's spiral screen is Class B and never raises a CRITICAL finding. Two further indications — dental panoramic and histopathology WSI — are built but deferred. The sensitivity figures below are validation-gate thresholds (pass/fail release criteria and design goals) — not measured, independently validated performance. Clinical validation runs ahead of each regulatory submission.
DICOM DX / CXR
An 11-finding vocabulary (pneumothorax, pneumonia/consolidation, TB, large effusion, severe cardiomegaly, mass/nodule, fracture and more). Critical findings are high-severity even at moderate confidence — a clinical-safety bias toward flagging. Target: ≥ 95% sensitivity on critical findings (validation gate, design goal — not yet independently measured), on our internal Indian-population test set.
DICOM MR
Three-way tumour classification plus WHO CNS grade (2-4) for gliomas. Suspected-malignancy triage with a target ≥ 92% sensitivity validation gate (design goal — not yet independently measured). Broader than an LVO-only stroke triage — this is a tumour-classification read.
DICOM MG
BI-RADS category with the 4A/4B/4C sub-category, laterality, and calcification-cluster detection across the CC + MLO bilateral views. Target: BI-RADS 4-5 sensitivity ≥ 88% (validation gate, design goal — not yet independently measured). A BI-RADS 4-5 read always routes to concurrence.
DICOM OCT
Diabetic-retinopathy ETDRS grade (none → proliferative), AMD status (dry / wet) and an optional glaucoma cup-to-disc ratio. Target: moderate-to-proliferative DR sensitivity ≥ 90% (validation gate, design goal — not yet independently measured).
Dermoscopy + Fitzpatrick
Melanoma probability plus a structured ABCDE dermoscopy assessment (asymmetry, border, colour, diameter, evolution), stratified by Fitzpatrick skin type I-VI for fairness monitoring. Target: melanoma sensitivity ≥ 87% (validation gate, design goal — not yet independently measured).
Spiral / wave drawing
Parkinsonism-likely flag plus a tremor-amplitude score from a spiral / wave drawing. Explicitly advisory — a screening aid, never a diagnosis, never a CRITICAL finding, never mandates concurrence. Target: sensitivity ≥ 82% (validation gate, design goal — not yet independently measured).
Deferred but built: dental panoramic X-ray (modalities/dental_xray.py, FDI tooth
numbering, Class B) and histopathology whole-slide imaging
(modalities/histopathology.py, ROI references, Class C). Every read carries a 1-5
Likert and a bias-stratum tuple — sex, age band, Fitzpatrick, scanner manufacturer, view
position — so every prediction logs its own subgroup stratification.
04 — Integration
MedSight follows one adapter pattern everywhere: a Protocol, a deterministic StubAdapter that CI exercises exclusively, and a ProductionAdapter slot filled at deploy time. So the test suite runs green with no PACS in the room, and the production wiring is a per-vendor config — not a rewrite. Production PACS / VNA / EMR connectors ship for Orthanc, dcm4chee, GE, Siemens, Philips, Hyland, Sectra, Mach7, Epic and Cerner / Oracle Health.
# Configure the connection to a hospital PACS. The production adapter # implements the same Protocol the StubAdapter does; CI runs the stub. from medsight.adapters.dicom.production import DicomNodeConfig, ProductionDicomAdapter cfg = DicomNodeConfig( calling_ae_title="MEDSIGHT_SCU", peer_ae_title="APOLLO_PACS", peer_host="10.0.4.20", peer_port=104, query_timeout_s=30.0, batch_size=25, retries=7, # exponential backoff, PRD Section 12 scp_ae_title="MEDSIGHT_SCP", scp_port=11112, # STORE SCP listener for pushed instances ) adapter = ProductionDicomAdapter(cfg) # Q/R a study, then de-identify at the edge before any cloud uplink. results = adapter.c_find(query) instances = adapter.c_move(study_uid, destination_aet="MEDSIGHT_SCP")
→ CLI surface: medsight init (generate appliance Ed25519 keypair) ·
medsight enroll <tenant-id> · medsight verify-chain (verify the
local audit chain end-to-end) · medsight deidentify (PS3.15) ·
medsight dossier <indication> --jurisdiction CDSCO_IN ·
medsight bsa-cert <indication> (seal a BSA §63 evidence bundle) ·
medsight modality-test <indication>.
05 — Audit-chain provenance
From 1 July 2024, every electronic record produced as evidence in an Indian court must satisfy Bharatiya Sakshya Adhiniyam §63 — hash, algorithm, chain of custody, device identity, operator. Medical-malpractice litigation, insurance-rejection challenges, NMC disciplinary proceedings and consumer-forum cases all turn on the technical integrity of the underlying record. MedSight produces a tamper-evident, §63-aligned evidence bundle engineered to satisfy those technical requirements — reusing the Ed25519-chained signer and per-tenant root-key custody pattern already shipping across the IcyCastle family. Admissibility in any proceeding is determined by the court.
Each entry carries the prior entry's hash and a 64-byte Ed25519 signature over the canonical payload. Editing any past entry breaks verification for every entry after it. Genesis links to 32 zero bytes.
The per-tenant Ed25519 root key lives in the HSM and is never persisted in application data — only its SHA-256 fingerprint is stored. Field appliances sign with an enrolled appliance key; production issuance uses the HSM-sealed root plus a distinct operator key.
medsight bsa-cert seals a cert PDF + signed JSON manifest over every chain entry into a tarball, with a detached signature sidecar that can be re-verified later. Single-study replay in ≤ 60 s; a 24-hour window in ≤ 5 min.
The regulatory-dossier exporter populates the clinical-validation, bias-evidence and provenance modules straight from the tenant's audit chain — per-event counts, the validation event span, and a reproducible anchor hash — so the submission tracks production reality, not a static snapshot.
Cross-border data-residency follows a closed compliance preset, not a feature flag: DPDP (India, Yotta Mumbai), HIPAA (45 CFR §164 — 6-year retention, 30-day DSAR, BAA-gated), GDPR, NHS-UK, PDPA-Singapore, Quebec25 and Australia-PR. The preset sets retention, audit verbosity, consent gates, de-identification depth and DSAR SLA in one toggle.
06 — Regulatory file · triple track (in progress)
MedSight is a triage aid: it never makes an autonomous diagnosis, and every high-stakes finding passes through two-radiologist concurrence before the EMR. That CADt posture is what keeps the software inside Class B / Class C — not autonomous Class D. The QMS, software-lifecycle and risk-management tracks feed one technical file that three regulators draw from. All three clearances and the medical-device-standard certifications below are on a regulatory track and not yet obtained — the file is design-conformant; submission and certification are targeted, in progress.
MDR 2017, Schedule MDR-3 (as amended G.S.R. 467(E))
Advisory indications are Class B; triage-influencing indications (CXR pneumothorax, mammography BI-RADS, brain MRI suspected-malignancy) are Class C. The dossier exporter auto-fills the audit-derived modules. First submission targeted for chest X-ray.
21 CFR 807.92 — 510(k) Premarket Notification
Predicate candidates under evaluation against FDA's cleared AI/ML SaMD database (cleared CADt chest-X-ray devices from Aidoc, Annalise.ai, Qure.ai and Lunit); the specific predicate and K-number are confirmed at submission. Software level-of-concern, cybersecurity and labelling per the guidance. HIPAA 45 CFR §164.312(b) audit controls addressed by the chain. Submission targeted (regulatory track, in progress).
Reg (EU) 2017/745 Rule 11 · Reg (EU) 2024/1689 Annex III §5
Rule 11 places triage SaMD in Class IIa, escalating to IIb where the outcome materially affects treatment urgency. The Ed25519 prediction log satisfies AI Act Art 12 (logging); two-radiologist concurrence satisfies Art 14 (human oversight). Notified body via BSI / TÜV SÜD / DEKRA.
IEC 62304:2006 +A1:2015 software lifecycle
Every software item carries a class: the modality triage pipelines and the concurrence / alert / audit-chain / key-custody items are Class C; transport adapters, scheduler and operator surfaces are Class B; the Parkinson's screen is Class B. The dossier manifest records the system class as C.
ISO 13485:2016 medical-device QMS
The QMS index maps the design-control, document-control and CAPA processes onto the existing IcyCastle ISO 27001 footprint. Certification via TÜV SÜD India targeted before the chest-X-ray CDSCO submission.
ISO 14971:2019 risk management
The risk file enumerates risks across data-acquisition, model-inference, integration, deployment, monitoring and decommissioning — each with severity, probability, detectability, residual risk after control, and traceability to the FR / NFR control that mitigates it.
Positioning is deliberate and compliant: assistive triage / CADt, never autonomous diagnosis. The radiologist is responsible for the final diagnosis (NMC Telemedicine Practice Guidelines 2020 §3.6); MedSight surfaces, prioritises and documents — it does not decide. Accreditation alignment: NABH 5th Edition IMS.7.b and JCI MOI.13.
07 — Multi-role clinician console
A radiologist does not see what a tenant admin sees; a read-only auditor cannot change a threshold; a federation grant needs two distinct operators. Every role assignment is itself anchored in the audit chain, so who-could-do-what at any point in time is reconstructable. The explanation overlay puts the Grad-CAM heatmap and per-finding Likert in front of the reading clinician.
Reads the study, sees the AI flag, the per-finding Likert and the Grad-CAM overlay, and provides the first concurrence sign-off. Can confirm, reject or modify a finding.
Provides the second, distinct concurrence sign-off on high-stakes findings and resolves discordance between two readers as the senior tie-breaker. The discordance is logged with a closed reason code.
Configures triage thresholds and reviews departmental performance. The CMIO owns the clinical-information posture and the worklist-priority policy across the site.
Sets the compliance preset, exports the regulatory dossier, and grants a cross-tenant federation agreement — which requires the closed-enum reason, an IEC-approval artefact, and two distinct attesting operators on each side.
Watches the per-site, per-indication, per-stratum bias-and-fairness metrics — sensitivity and specificity by sex, age band, Fitzpatrick type, scanner manufacturer and view position — with warning / critical drift flags.
A read-only auditor verifies the chain and the dossier but changes nothing. The patient role is DSAR-only: access, erasure, amendment or export, with identity verification by Aadhaar OTP, photo ID or passport, severing the pseudonym linkage on erasure.
08 — Deployment + capability
MedSight runs the model on a hospital-LAN edge appliance and keeps the multi-tenant control plane in in-country data centres (CERT-In-empanelled hosting targeted, in progress). SQLite with WAL today (M0/M1), schema-portable to Postgres at M2 — every tenant in its own database, isolation enforced at every query. Federated learning is available to chain customers without moving raw scans off-site.
275 TOPS, 60 W TDP, India-assembled. CXR inference clears in ≤ 5 s p99; the appliance runs offline with an audit-chain-buffered cache and a heartbeat beacon to the control plane.
A Dell PowerEdge R7625 with dual NVIDIA L40S for diagnostic-chain volumes, or an x86 + Hailo-15H variant. NDAA Section 889-clean BOM track available alongside the default track.
Yotta Mumbai for DPDP tenants; AWS Mumbai / Singapore / US-East for HIPAA-tagged tenants. Per-tenant database files, AES-256 at rest, TLS 1.3 in transit, and a target ≥ 99.95% cloud availability per quarter (design goal SLA, not yet independently measured).
A federation agreement lets a chain improve a model across sites under a closed-enum reason, an ethics-committee approval artefact and a two-operator gate on each side — without raw patient scans leaving any site. Revocation is one call; every query is audit-signed both ways.
09 — How to engage
MedSight is available in early access for hospitals, diagnostic chains, hospital groups and insurers. Two engagement tracks are open now, pre-clearance: a research MoU (AIIMS / state-medical-college, zero-cost — training-data access and a co-authored validation paper in exchange for a pilot deployment) and a design-partner track for chains that want to deploy and co-validate ahead of CDSCO clearance. Two-radiologist concurrence is in the base licence, not an add-on. Reach out to discuss fit, deployment scope and regulatory timing.
10 — Start
Assistive triage across six study types, two-radiologist concurrence built in, and an Ed25519 audit chain engineered to produce a tamper-evident, BSA §63-aligned evidence bundle. 760 software tests (pipeline / integration — not a measure of diagnostic accuracy) — request early access and we'll run the suite and the BSA §63 bundle live. Reference implementation, pilot-ready; admissibility in any proceeding is determined by the court.