Cloud-Based Integrated Diagnostics Platform with AI-Assisted Digital Pathology

philips · Healthcare · 2024 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.5 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Cloud-native integrated diagnostics and digital pathology

Philips needed to move large, siloed medical imaging workflows into a cloud-native integrated diagnostics platform so clinicians could access radiology, cardiology, and pathology data together, collaborate across regions, and apply AI to pathology workflows at clinical scale.

Constraints:

Architecture

Philips built a cloud-native enterprise pathology and integrated diagnostics platform on AWS. High-performance scanners generate DICOM whole-slide images, staged data lands in S3 buckets, AWS HealthImaging ingests and indexes image sets and tiles, hot/mid/archive lifecycle management controls cost and latency, stateless services run on containers/Lambda with Aurora Serverless metadata, AI algorithms perform quality control and measurement, and zero-footprint viewers let pathologists review, bookmark, report, collaborate, and connect data to SageMaker for model training and deployment.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  high-performance-scanner["high performance scanner"] --> s3-staging-buckets["S3 staging buckets"]
  s3-staging-buckets --> aws-healthimaging["AWS HealthImaging"]
  aws-healthimaging --> image-set-index["image set index"]
  aws-healthimaging --> tile-retrieval-apis["tile retrieval APIs"]
  aws-healthimaging --> storage-lifecycle-manager["storage lifecycle manager"]
  aws-healthimaging --> ai-assisted-pathology-algorithms["AI-assisted pathology algorithms"]
  ai-assisted-pathology-algorithms --> pathology-workflow-engine["pathology workflow engine"]
  pathology-workflow-engine --> zero-footprint-viewer["zero-footprint viewer"]
  pathology-workflow-engine --> laboratory-and-ehr-integrations["laboratory and EHR integrations"]
  tile-retrieval-apis --> zero-footprint-viewer
  image-set-index --> integrated-diagnostics-workspace["integrated diagnostics workspace"]
  zero-footprint-viewer --> collaboration-layer["collaboration layer"]
  integrated-diagnostics-workspace --> collaboration-layer
  ai-assisted-pathology-algorithms --> sagemaker-training-deployment-loop["SageMaker training deployment loop"]
Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component Role Interacts with
high performance scanner Digitizes pathology slides into DICOM whole-slide images. S3 staging buckets
S3 staging buckets Stage large pathology data in cloud storage before HealthImaging ingestion. high performance scanner, AWS HealthImaging
AWS HealthImaging Ingests, stores, indexes, tiers, and serves medical images and pathology tiles. S3 staging buckets, image set index, tile retrieval APIs, storage lifecycle manager
image set index Groups related medical images and metadata into searchable AWS resources. AWS HealthImaging, integrated diagnostics workspace
AI-assisted pathology algorithms Perform automatic quality control, measurement, pre-diagnosis, and categorization before pathologist review. AWS HealthImaging, pathology workflow engine, SageMaker training deployment loop
pathology workflow engine Moves slides through quality control, worklists, review, reporting, and rules-based next steps. AI-assisted pathology algorithms, zero-footprint viewer, laboratory and EHR integrations
zero-footprint viewer Web-based viewer for pathologists to navigate whole-slide tiles without local software. tile retrieval APIs, pathology workflow engine, collaboration layer
tile retrieval APIs Fetch requested whole-slide tiles with low latency for web viewing. AWS HealthImaging, zero-footprint viewer
storage lifecycle manager Moves pathology data across hot, mid-tier, S3, AHI, and Glacier archive layers. AWS HealthImaging
integrated diagnostics workspace Shows radiology, pathology, and cardiology data together with shared metadata. image set index, collaboration layer
collaboration layer Enables regional collaboration, tumor boards, bookmarks, notes, and consensus review. zero-footprint viewer, integrated diagnostics workspace
SageMaker training deployment loop Uses centralized imaging data to train, specialize, evaluate, and redeploy AI models. AI-assisted pathology algorithms
laboratory and EHR integrations Connect pathology workflows to lab information systems and electronic health records. pathology workflow engine

Data flow

A pathology slide is scanned into a DICOM whole-slide image, staged in cloud storage, consumed by AWS HealthImaging, indexed as image sets and tiles, enriched with metadata, processed by AI-assisted quality-control and measurement algorithms, displayed through a web viewer with sub-200 ms tile retrieval, linked to notes/bookmarks/reports, and made available for collaboration, tumor boards, and downstream SageMaker model training/deployment.

4 supporting quotes
“error-prone but also not very efficient. Digital pathology changes this in a very fundamental way. The moment the slide is smeared and ready, it goes into a scanner. These are high performance scanners. And the moment the slide is digitized, it becomes part of the cloud-native imaging”transcript, stated
“the moment the slide is digitized, it becomes part of the cloud-native imaging fabric. It's enriching metadata. We have AI-assisted algorithms that start the automatic measurements and measurements taking. There are other Philips and third party algorithms”transcript, stated
“We stream the data directly in these three buckets, from which AHI actually consume the data. If you're wondering why the S3 bucket in between when AHI actually has a DICOMweb interface, well, for the size of data of pathology, it's actually better to stage”transcript, stated
“We can request any slide, any tile from any slide, and get a sub-200 millisecond response. That means that the pathologist can literally click through a slide at any speed and never notice any latency of images loading. The other part is, and again, I'll go”transcript, stated

Why this architecture

Philips uses cloud-native imaging because pathology files are too large for traditional on-prem and siloed workflows: central cloud storage, HealthImaging tile indexing, web viewers, and stateless services allow fast access, collaboration, lifecycle management, and AI training without copying data into separate research silos.

3 supporting quotes
“There was simply not proper architecture and proper resources on premise in hospital to build a solution like this.”transcript, stated
“With integrated diagnostic, the data is in one place and not just the data, but the metadata is in one place. It is all connected, it is all well-described. It is all predefined. So, things like training multimodal AI models becomes much easier because the”transcript, stated
“Connecting that with other systems such as SageMaker, for example, allows us to bring that data, train or specialize a model”transcript, stated

Models used: AI-assisted pathology algorithms · custom medical imaging models · SageMaker-trained models · AWS HealthScribe

Scale

Alternatives considered (roads not taken)

Manual microscope and physical-slide workflow

Rejected because: Manual workflow requires moving slides and notes and takes far longer from slide to report.

“Instead of zooming in and out on a microscope, they're looking at a screen of a high-resolution image. Thanks to AHI, they're able to pull in slides, focus on specific areas, and easily switch between different areas of the slide. Not only that, they can”transcript, stated
“manual movement of the slide, manual notes. You have to gather a bunch of slides to move them forward. In digital pathology, this goes down to 36 minutes. This is a very important time save, not just for pure sake of efficiency, but this means that now,”transcript, stated

Siloed radiology, cardiology, and pathology systems

Rejected because: Separate systems require screenshots, printouts, offline sharing, and point-to-point integrations instead of unified multimodal review.

“you are dependent on either operating different systems, taking screenshots, either doing printouts, or sharing data offline”transcript, stated
“it's much more easy to operate. And finally, data, and this is the most underestimated thing. Today, data lives in silos. Radiology data lives in the packs. Cardiology data lives in the cardiology solution. Pathology data, if it's digitized,”transcript, stated

Keep all pathology data in hot storage

Rejected because: Pathology data is too large and unpredictably accessed, so lifecycle tiering is required for cost and availability.

“If you try to store them all in a very hot storage, it's very difficult. They're large, it gets expensive. AHI removes that problem for us. It adds an obstruction layer in which we can request whatever we want and we always get it with a specific amount of”transcript, stated
“smartly move the data to different storage layers such as S3, AHI, and then Glacier for archiving. On the clinical diagnosis part, and this is also extremely important, because of the integration of AHI, like I said, with things like SDS in EIM, we can have”transcript, stated

Train AI models by copying data into separate locations

Rejected because: Centralized cloud imaging plus SageMaker lets teams train, evaluate, and redeploy models without copying data into multiple separate locations.

“Having the data in one place and not being needing to copy it in multiple separate location is what distinguishes AWS HSI for us.”transcript, stated

Patterns

latency-optimization (clinical-viewing)

AWS HealthImaging indexes tiles and provides sub-200 ms tile retrieval so pathologists can navigate huge slides without noticeable loading.

“We can request any slide, any tile from any slide, and get a sub-200 millisecond response. That means that the pathologist can literally click through a slide at any speed and never notice any latency of images loading.”transcript, stated
“That means that the pathologist can literally click through a slide at any speed and never notice any latency of images loading. The other part is, and again, I'll go here on the lifecycle management of things.”transcript, stated

batch-offline-processing (medical-imaging-ai)

Pathology images can be used to train, specialize, evaluate, and redeploy AI models through SageMaker.

“allows us to bring that data, train or specialize a model, evaluate that model, and then deploy it back in the solution”transcript, stated

data-quality-loop (clinical-data)

The system keeps data and metadata connected in one platform so imaging, reports, bookmarks, and model training all refer to the same source of truth.

“the data is in one place and not just the data, but the metadata is in one place.”transcript, stated
“Thanks to the digitization of images, we always looking at the same source of truth. And last but not least is research.”transcript, stated

observability-tracing (clinical-workflow)

Bookmarks, reports, metadata, and slide links make diagnostic claims traceable back to the source image region.

“thanks to bookmarks, finding that report can be linked directly in the slide.”transcript, stated
“So, when you're reading the report, because it is all running on the same platform, you can actually click on different bookmarks and directly zoom back to the source of truth. And finally, efficiency.”transcript, stated

human-in-the-loop-gate (clinical-ai)

AI assists technicians and pathologists with quality control, measurements, and categorization, while clinical diagnosis remains in the pathologist workflow.

“Quality control becomes verification of previously done AI-assisted control.”transcript, stated
“to help the technician and then the pathologist following that to actually perform their work.”transcript, stated

Tradeoffs

Stage pathology data in S3 before HealthImaging ingestion.

Chose Cloud staging buckets over Direct DICOMweb ingestion only.

Rationale: For very large pathology data, staging makes transfer corruption or metadata fixes easier to diagnose and repair.

“for the size of data of pathology, it's actually better to stage the data in the cloud first.”transcript, stated
“re-upload it if it only requires, for example, a metadata re-upload or a metadata fix.”transcript, stated

Use lifecycle tiering instead of keeping everything hot.

Chose Hot, mid-tier, S3/AHI, and Glacier-style storage workflows over All-hot storage for all pathology images.

Rationale: Pathology data must remain available for future comparisons, but holding all historical slides in hot storage is impractical.

“data is always important, but holding that amount of data in hot storage is practically impossible.”transcript, stated
“it's moved to proper tiers of storage. It also, thanks to other AWS services like Glacier, we can also decide to move this data into an archival storage. Because the thing with pathology data is, it's important when it's hot, when a diagnosis is made, but it also becomes important at points in the”transcript, stated

Use zero-footprint cloud viewers.

Chose Web-based viewers fetching images directly from AHI over Installed desktop pathology clients.

Rationale: Cloud viewing enables real-time regional collaboration and avoids local software installation.

“zero footprint viewers, meaning clients don't have to install software. It's all web-based.”transcript, stated
“it allows real time access to that data to multiple pathologists so they can collaborate in real time.”transcript, stated

Gotchas (what bit them)

Pathology data breaks naive image-storage assumptions.

Root cause: Whole-slide images are huge, tile-rich, and need metadata-linked access at micron resolution.

Resolution: Use HealthImaging image sets, tile APIs, staging, metadata, and lifecycle management instead of treating slides as ordinary files.

“It's gigapixel images. So, you're imaging cancer tissue, tissue specimen, mounting it on a glass slide, and imaging this data at micron resolution.”transcript, stated
“How do you get specific data contained within that vast image to a client for a physician to review”transcript, stated

Data corruption and metadata issues can happen during scanning/transfer.

Root cause: Large pathology uploads may fail partially or require metadata repair.

Resolution: Stage data in cloud buckets before HealthImaging ingestion so teams can inspect and fix transfer issues.

“Because data can get corrupted during transfer. Metadata can be changed, metadata can be corrupted.”transcript, stated
“it's better for us to have the data in one place where we can figure out what happened”transcript, stated

Siloed diagnostics force clinicians into unsafe manual coordination.

Root cause: Radiology, cardiology, and pathology data live in different systems with different workflows and integration standards.

Resolution: Use one integrated diagnostic workspace and shared data/metadata layer.

“Today, that is all either a very cumbersome and error-prone point to point integration or a manual step”transcript, stated
“different upgrades of software, different database upgrades, different standards, different networking requirements, now you're going to work with one integrated solution that is much more easy to manage and it's much more easy to operate. And finally, data, and this is the most underestimated”transcript, stated

Evaluation

Philips evaluates the platform through workflow time, image-ingestion scale, tile latency, pathologist survey results, collaboration, and diagnostic consensus rather than LLM-style text benchmarks.

Metrics: 11 hours 35 minutes manual slide-to-report time · 36 minutes digital pathology slide-to-report time · sub-200 ms tile retrieval · under one minute whole-slide image generation · almost one gigabyte per minute per scanner · 10+ scanners at some hospitals · 52-pathologist survey · 21% more cases diagnosed digitally

3 supporting quotes
“It's about 11 hours and 35 minutes from slide to report. A lot of this is because there is a map manual movement of the slide, manual notes.”transcript, stated
“We can request any slide, any tile from any slide, and get a sub-200 millisecond response. That means that the pathologist can literally click through a slide at any speed and never notice any latency of images loading.”transcript, stated
“And this is based on a survey of 52 pathologists, lab agents, and things like that. And I'm not gonna go through the whole thing.”transcript, stated

Outcomes

The platform addresses petabyte/billion-image scale, four-gigabyte pathology files with 100,000+ images, sub-200 ms tile retrieval, almost one gigabyte per minute per scanner, 10+ scanners in some hospitals, and 52 surveyed pathologists.

3 supporting quotes
“There is about 21% more cases diagnosed in digital versus manual pathology.”transcript, stated
“100% of the surveyed pathologists do not want to go back to a microscope once they experience digital pathology.”transcript, stated
“100% of pathologists have claimed in this survey expressed that digital pathology helps them reach diagnostic consensus”transcript, stated