Annalise.ai

Deep learning detection across two modalities is the entire product. The enterprise chest X-ray product detects up to 124 findings in under 20 seconds and the head CT product covers up to 130 radiological findings on non contrast studies, which is comprehensive coverage rather than the single indication approach most imaging AI takes. The company has also built a foundation model applied to report generation, positioning the same underlying models beneath both detection and drafting.

Two distinct products with different oversight postures, and buyers should not conflate them. The FDA cleared portfolio is triage and notification, flagging suspected time critical pathology and prioritizing cases within the radiology worklist so the care team treats the most critical patients first, with the radiologist reading every study. The comprehensive enterprise detection product covering 124 or 130 findings operates as clinical decision support alongside interpretation. Neither renders a diagnosis. The report generation work is the direction where oversight questions will sharpen, and no published detail describes how drafted findings are reviewed.

Training data provenance is described with unusual specificity for this category: models are stated to be trained on some of the world's largest and most diverse radiologist hand annotated datasets, and the company names hand annotation rather than leaving labeling methodology vague.

A published accuracy figure exists for the head CT product, reported as a 32 percent improvement in diagnostic accuracy averaged across analyzed findings, though the comparator and study design are not detailed in the materials reviewed. Model architecture and per finding performance are not published.

The training corpus is described in more detail than almost anything else in this index, which answers the data half of this axis properly and makes the remaining question specific rather than general. Source composition is stated across three continents, multiple equipment manufacturers, departmental and portable imaging, varied demographics and both inpatient and outpatient populations, with an independent three radiologist labelling procedure.

A reader can judge whether that corpus resembles their own population, which is the point of the disclosure. What is unaddressed is the customer side, and the company's own description is what makes it pressing. An organisation that has assembled a corpus of that size, and continues to process imaging at national scale in more than one health system, has an evident and legitimate interest in the studies flowing through its products.

Nothing published states whether studies processed for customers feed model development, whether that is severable in contract, what retention applies to imaging or derived findings, or what is returned or deleted at termination. Jurisdictional spread compounds it, since the answer may differ by market and no market specific terms are published. Nothing else is enumerated either, with no hosting arrangement or sub processor list located. Ask whether processed studies train models, whether a site can decline, and for retention and deletion terms per market.

Deployment scale is the strongest evidence and it is substantial: clearance in more than 40 countries with deployment in 15, accessibility to half of radiologists in Australia, and use in processing more than 35 percent of chest X-rays in England, which is national infrastructure scale rather than pilot deployment. The company also reports strong results in an independent United States healthcare AI challenge for report generation.

Regulatory validation spans ten separately cleared findings. What is thinner is peer reviewed outcome evidence showing changed patient results, and the 32 percent accuracy improvement claim lacks published study detail.

Converted from Not Rated. No stewardship terms were published, and the company's own description of its training data makes the unanswered question a specific one rather than a general one.

What is disclosed about data is unusually good, and it sits on the model side rather than the stewardship side. The chest radiograph model is described as trained on more than 750,000 images sourced across three continents, spanning different equipment manufacturers, departmental and portable imaging, varied patient demographics, and both inpatient and outpatient groups, with cases hand labelled independently by three radiologists under standardised labelling procedures. Describing the composition of a training set at that level of detail is rare in this index and it is creditable.

It also raises the question this axis exists to answer. A company that has assembled a corpus of that size, and continues to process imaging at national scale in more than one health system, has an evident interest in customer data. Nothing published states whether studies processed for customers feed model development, whether that is severable in contract, what retention applies to imaging or derived findings, or what is returned or deleted at termination.

The jurisdictional spread compounds it, since the answer may differ by market and no market specific terms are published.

Ask whether processed studies train models, and whether a site can decline.

Converted from Not Rated. No published position was located, and the corporate structure makes it a question a United States buyer must settle explicitly.

No business associate agreement, addendum, role statement, subcontractor flow down, breach notification timetable or review cadence was retrieved.

The role itself is not in doubt. A hospital or imaging provider is the covered entity; a vendor receiving studies and returning findings processes protected health information on its behalf and is a business associate with direct liability.

The structure is what needs resolving. The company is an Australian headquartered joint venture with a separate United States presence and cleared products in the American market, and it operates across more than forty jurisdictions with a significant footprint in publicly funded health services elsewhere. Which legal entity contracts, whether any processing, model operations or support access occurs outside the United States, and how affiliate access is controlled are all live questions that only the agreement answers. Where support or engineering touches imaging from another country, the subcontractor terms are the sole control.

The edge deployment option is worth raising in the same conversation, since a local installation narrows what the vendor receives at all and therefore what the agreement has to cover.

Ask which entity signs, whether access occurs outside the country of care, and for the subcontractor terms in writing.

Converted from Not Rated. No enterprise attestation was located across two differently phrased searches, including one using the company's own product names.

No SOC 2 report of either type, no HITRUST certification, no ISO 27001, no trust centre, no penetration testing statement and no subprocessor disclosure were retrieved. Absence of a retrieved document is not proof none exists, and an organisation deployed across national health services at this scale will have satisfied institutional security review many times. None of it is public.

One thing does exist and it should be described accurately rather than credited as more than it is. The products are regulated as software as a medical device and are approved for sale in more than forty countries, which means cybersecurity documentation forms part of regulatory submissions and is examined by reviewing authorities. That is a genuine external examination and this index has credited it elsewhere in imaging. It examines the security of the product. It is not an audit of the company's enterprise information security posture, which is what a buyer's vendor risk process is asking about, and it is not a substitute for one.

The comparison within this index is direct: a peer in the same category holds ISO 27001 with periodic external surveillance audits alongside the same regulatory route, and graded higher for it. That combination is achievable here and is not evidenced.

Ask what independent assessment the company itself has undergone, and for the report.

Ten FDA cleared findings in the United States, split five for chest X-ray and five for head CT, built through successive 510(k) clearances beginning February 2022 with pneumothorax triage. The chest portfolio was reported as the highest number of cleared triage findings on chest X-ray in the US market, and the first product to differentiate tension pneumothorax so sites can apply distinct triage rules to the most urgent cases, which is a genuinely useful clinical distinction rather than a marketing one.

The company also received the first Breakthrough Device Designation ever granted to a radiology triage device, for obstructive hydrocephalus. Regulatory reach extends to clearance in more than 40 countries.

No formal governance framework was located, but the emphasis on large and diverse hand annotated training datasets, combined with clearance across more than 40 countries and deployment at national scale in health systems with different populations and equipment, implies validation across substantial real world variation. No published subgroup performance analysis addresses demographics or acquisition equipment directly.

How the models were built is described unusually well and how they perform is described in a form that hides what matters. On construction the disclosure is genuine: the chest radiograph model is stated to be trained on more than seven hundred and fifty thousand images sourced across three continents, spanning different equipment manufacturers, departmental and portable imaging, varied patient demographics and both inpatient and outpatient groups, with cases hand labelled independently by three radiologists under standardised procedures.

Naming hand annotation rather than leaving labelling vague, and describing the breadth of the sources, tells a reader why the model might generalise, which is the right argument to make in imaging. The performance claim does the opposite.

A thirty two per cent improvement in diagnostic accuracy averaged across analysed findings conceals exactly the variance that matters: these models cover many findings, the findings differ enormously in difficulty and in consequence, and an average lets a strong result on common obvious findings carry a weak one on the rare subtle finding a radiologist most needs help with.

The comparator and study design are not detailed either, so the baseline the improvement is measured against is unknown. No warranty, indemnity or remediation commitment was located. Ask for per finding sensitivity and specificity, and what the thirty two per cent was measured against.

Products deploy into radiology worklists and reading workflows, with triage findings flagged for prioritization inside existing worklist tooling, which is the necessary integration surface. No named PACS integrations, standards support, or platform partnerships were enumerated in the materials reviewed.

Converted from Not Rated. Two deployment architectures exist and both are documented, which is more than the earlier review found.

The enterprise products run in the cloud, and Amazon Web Services is named as the underlying infrastructure in deployment reporting from at least one large hospital group. Alongside them sits a separate edge product, a point of care chest radiograph solution returning findings in about ten seconds. An edge deployment means processing happens locally rather than in a remote environment, which is architecturally the strongest available answer to a data location question.

One honest qualification. The company markets the edge product for speed at the point of care rather than as a residency control, and nothing published states what leaves the device or whether results, images or telemetry are transmitted onward. A buyer with localisation obligations should confirm that directly rather than assume the architecture settles it.

What is absent is everything else a residency review asks. No region list, no tenancy model, no statement of whether customer environments are separated, no subprocessor list and no retention terms for imaging. That gap is more consequential here than for a single market vendor, because the products are approved for sale in more than forty countries whose data localisation rules differ materially, and the company itself notes that product availability varies by market.

Ask which architecture your site would run, in which region, and what the edge product transmits.

No published pricing, but one materially useful reimbursement fact is public: the company holds Medicare New Technology Add on Payment status in the United States, which provides a defined reimbursement pathway and changes the return calculation for a hospital buyer. Knowing NTAP applies is more actionable than most vendors in this index offer. Product structure spans separately licensed triage and comprehensive detection products, and neither is priced publicly.

Broad on both modality and finding count. Coverage spans chest radiography at up to 124 findings and non contrast head CT at up to 130 findings, with US cleared triage across pulmonary and neurocritical indications including a suite of time sensitive stroke findings. Settings span emergency, inpatient, and outpatient reading, and geographic reach covers clearance in 40 plus countries with deployment in 15, including national scale use in a public health system. Few vendors in this index combine per modality depth with this geographic footprint.