## Overview This case study presents Philips' comprehensive cloud transformation journey in partnership with AWS, focusing on building an integrated diagnostics platform that leverages AI and machine learning for medical imaging, particularly in digital pathology. While the presentation primarily discusses medical imaging infrastructure and digital pathology workflows, it touches on AI/ML deployment aspects relevant to LLMOps, including the use of AWS services for training, deploying, and operationalizing AI models in production healthcare environments. The case study also mentions AWS HealthScribe for generating clinical notes from doctor-patient interactions using generative AI, though this is not the primary focus. Philips, a major healthcare technology company serving 95 of the top 100 US hospitals, embarked on an ambitious initiative to migrate their entire healthcare informatics portfolio to the cloud. The partnership with AWS spans nearly two decades since 2008, with Philips being the first company to launch production medical workloads to the cloud in 2014-2015. Over the past three years since forming their Enterprise Informatics business, Philips has aggressively shifted their portfolio to cloud-native architectures, launching their Health Suite imaging cloud PACS system in 2023 with over 150 customer sites migrated, and recently announcing their cardiovascular workspace in the cloud. ## The Problem Space Healthcare faces multiple systemic challenges that impact the deployment of AI and ML solutions. Rising costs of care, aging populations, complex disease states, and staffing shortages create immense pressure on healthcare systems. A critical insight from the Philips Future Health Index survey reveals that 77% of healthcare professionals across nearly 2,000 respondents globally report a near 10% loss of productivity simply due to inability to access the right data. This data accessibility problem is compounded by disparate data silos created during the transition from paper-based to electronic systems, where different specialties (radiology, pathology, cardiology, laboratory, EMR systems) operate isolated solutions that don't communicate effectively. Medical imaging data presents unique technical challenges for AI/ML deployment. The data is multi-modal with different characteristics: radiology studies might contain hundreds of images per patient, pathology whole slide images can be 4+ gigabytes containing 100,000+ frames at gigapixel resolution, and cardiology generates waveform data alongside imaging. The DICOM (Digital Imaging and Communications in Medicine) standard, while ubiquitous, is 30-40 years old and was originally designed as a network protocol rather than a storage standard, creating complexity when co-opted for data storage. Each DICOM file contains both pixel data and extensive metadata (hundreds of thousands of potential attributes), and the resource hierarchy requires sophisticated querying to locate specific patient studies across potentially billions of objects and petabytes of data. ## Technical Architecture and Infrastructure Philips' solution architecture leverages AWS Health Imaging (AHI) as the foundational managed service to eliminate undifferentiated heavy lifting. Traditional DIY architectures for medical imaging require complex combinations of S3 for storage, Lambda and EventBridge for ingestion triggers, RDS for relational metadata, OpenSearch for semantic search across patient records, Step Functions for workflow orchestration, Fargate for data processing and transcoding legacy compression formats, and EC2/API Gateway for API interfaces. AWS Health Imaging consolidates this complexity into a purpose-built managed DICOM store that handles metadata parsing, pixel encoding/decoding, DICOMweb interfaces, image frame indexing with unique IDs, and automated data lifecycle management across storage tiers. The architectural approach for Philips' digital pathology solution demonstrates cloud-native design principles relevant to ML operations. High-performance scanners generate whole slide images at approximately 1 gigabyte per minute per scanner, with hospitals running 10+ scanners and potentially thousands of hospitals in the ecosystem. Rather than streaming directly to AWS Health Imaging's DICOMweb interface, Philips stages data in S3 buckets first, allowing for data integrity verification and metadata correction before ingestion into AHI. This staging approach prevents corruption issues and enables retry logic without re-uploading from hardware. AHI then consumes from S3, automatically organizing data into patient-centric image sets, normalizing metadata into searchable JSON structures with versioning for immutable audit trails, and managing hot/warm/cold storage tiers transparently. The solution is designed as fully stateless and horizontally scalable. Most application logic runs in containers, with AWS Lambda handling event-driven workflows. Aurora Serverless manages patient metadata storage. This stateless architecture enables virtually endless scalability, critical for handling growing patient volumes and data sizes. The system integrates with AWS Glacier for long-term archival of pathology data, which remains diagnostically relevant for years when evaluating tumor progression but doesn't require constant hot storage access. ## AI/ML Integration and LLMOps Considerations While the presentation focuses primarily on infrastructure, several aspects relate to AI/ML operations in production. Philips implements AI-assisted algorithms throughout the digital pathology workflow: automatic quality control of slide preparation, automatic measurements and categorization, pre-diagnosis assistance, and image processing for dyeing and feature identification. These algorithms run as part of the ingestion pipeline, enriching slides with metadata and AI-generated insights before reaching pathologists. The sub-200 millisecond tile retrieval performance from AWS Health Imaging enables real-time AI inference, as models can rapidly access any region of gigapixel pathology images without latency. The architecture supports ML model training and deployment workflows through integration with Amazon SageMaker. Philips emphasizes that having data consolidated in AWS Health Imaging rather than scattered across multiple systems enables researchers to train multi-modal AI models without copying petabytes of data to separate research environments. The unified data fabric allows connecting pathology images with radiology scans, cardiology data, and clinical records for training models that leverage multiple diagnostic modalities. Once models are trained and validated in SageMaker, they can be deployed back into the production pathology solution without complex data movement or management overhead. The presentation mentions AWS HealthScribe as part of the broader AWS healthcare portfolio for "generating notes from doctor-patient interactions," indicating use of generative AI for clinical documentation. This represents a direct LLM application in production healthcare workflows, though the case study doesn't elaborate on implementation details. The mention suggests Philips envisions generative AI as part of their integrated diagnostics vision, potentially for synthesizing insights across multiple diagnostic modalities or assisting with report generation. ## Clinical Workflow Transformation The digital pathology solution demonstrates how cloud infrastructure and AI can fundamentally transform clinical workflows. Traditional manual pathology involves physical glass slides, manual labeling, quality control inspection under microscope, pathologists examining slides through eyepieces while taking separate notes, and finally writing reports disconnected from the source images. This process averages 11 hours and 35 minutes from slide to final report, according to surveys cited by Philips. With Philips' cloud-native digital pathology, slides are digitized immediately after preparation and enter the AI-enhanced workflow. Automated algorithms perform initial quality control and measurements, pathologists review high-resolution digital images on screens with AI assistance for region identification, bookmarks and annotations are created directly on digital slides, and reports are written in parallel while viewing images with hyperlinked bookmarks connecting report findings to specific slide regions. This reduces the slide-to-report time to 36 minutes on average—a dramatic improvement that increases pathologist capacity to handle more cases and reduces time to diagnosis for patients. The cloud-based approach enables zero-footprint web viewers that access images directly from AWS Health Imaging through secure APIs with IAM integration. Pathologists don't install software but access the system through browsers. This enables real-time collaboration where multiple pathologists can simultaneously view and annotate the same slide, regardless of geographic location. A pathologist in the US can consult with a specialist in the Netherlands by simply sharing access to the digital slide, whereas traditional pathology would require physically shipping glass slides internationally. This collaboration capability extends to tumor boards, where multidisciplinary teams meet to discuss complex cases—previously offline meetings with printed materials now become fully digital sessions with live access to all imaging modalities and the ability to dynamically explore slides based on discussion. ## Integrated Diagnostics Vision Philips' broader vision of "integrated diagnostics" aims to break down specialty silos and provide clinicians with unified access to all patient imaging data through a single interface. Rather than radiologists, cardiologists, and pathologists working in separate systems with point-to-point integrations, the cloud platform creates a longitudinal patient view where any specialist can access relevant multi-modal data. Pathology data serves as a critical integrative element—70% of imaging decisions are made based on pathology findings, and 100% of cancer diagnoses require pathology confirmation, yet pathology was the last imaging modality to digitize due to technical challenges. The integrated approach has significant implications for AI/ML deployment. Multi-modal models that combine radiology, pathology, and clinical data for improved diagnosis or treatment planning become feasible when data lives in a unified cloud fabric rather than isolated on-premises systems. The consolidated metadata and common access patterns across modalities simplify the development of AI workflows that span specialties. Philips emphasizes that with data "in one place" rather than copied across multiple research environments, training sophisticated cross-modality models becomes "much easier because the access to the data is there." ## Scale and Impact The production deployment operates at significant scale. Philips has migrated 134 petabytes of medical imaging data to AWS, encompassing 34 million patient exams and 11 billion medical records and images. Over 150 customer sites have migrated to the cloud PACS system. The infrastructure supports 82% of the top 250 US hospitals using Philips healthcare informatics solutions across radiology, cardiology, and pathology specialties. According to survey data from 52 pathologists and lab managers cited in the presentation, 100% of pathologists who experienced digital pathology "do not want to go back to a microscope," indicating strong clinical adoption. The same survey reports 21% more cases diagnosed in digital versus manual pathology, suggesting productivity improvements. All surveyed pathologists (100%) indicated that digital pathology helps them reach diagnostic consensus through enabled digital collaboration. ## Operational Considerations and Tradeoffs The case study presents a highly positive view of the cloud migration and AWS Health Imaging benefits, which should be balanced with recognition of inherent tradeoffs. The presentation emphasizes removing "undifferentiated heavy lifting" through managed services, but this comes with vendor lock-in to AWS-specific services like Health Imaging that don't have direct equivalents in other clouds or on-premises environments. Organizations must weigh the productivity gains against strategic flexibility. Data staging in S3 before ingestion to Health Imaging adds latency and complexity compared to direct streaming, though Philips justifies this for data integrity verification. The sub-200 millisecond tile retrieval performance is presented as enabling real-time pathology review, but network latency variations across geographies aren't discussed—a consideration for international deployments. The presentation doesn't detail model monitoring, versioning, or governance for the AI algorithms deployed in production workflows. How are AI-assisted quality control and measurement algorithms validated? How are model updates deployed without disrupting clinical workflows? What mechanisms ensure AI recommendations are appropriately calibrated and don't introduce bias? These LLMOps concerns are relevant but not addressed in the presentation. Cost optimization strategies beyond general storage tiering aren't explored. While Health Imaging manages hot/warm/cold data automatically, the presentation doesn't quantify storage cost savings or discuss compute costs for AI inference at scale. Healthcare organizations considering similar architectures would need detailed cost modeling. ## Partnership and Ecosystem The case study strongly emphasizes the partnership model between Philips and AWS spanning nearly two decades. Philips positions this as essential for success in the complex, highly regulated healthcare environment. Philips contributes clinical expertise, deep understanding of regulatory requirements, established relationships with hospital systems, and domain knowledge of medical imaging modalities. AWS provides secure, scalable cloud infrastructure, purpose-built healthcare services (Health Imaging, HealthLake, Health Omics, HealthScribe), continuous innovation on underlying services, and expertise in cloud-native architecture patterns. This partnership approach enabled Philips to be first-to-market with production medical workloads in the cloud in 2014-2015 and to aggressively migrate their entire portfolio over the past three years. The co-innovation model includes working directly with AWS product teams to shape services like Health Imaging based on real-world pathology requirements—for example, the need to index and rapidly retrieve individual tiles from gigapixel images containing 100,000+ frames. The presentation positions AWS Health Imaging as removing complexity so Philips can "focus on what we are good at"—clinical innovation, user experience design, and differentiated AI capabilities rather than building and operating storage infrastructure. This represents a strategic choice to leverage managed services for foundational capabilities while concentrating engineering resources on higher-value features and clinical workflows. ## Conclusions and Broader Implications This case study illustrates how cloud infrastructure and managed services enable AI/ML deployment at scale in highly regulated, mission-critical healthcare environments. The success factors include starting with solid data foundations (unified storage, normalized metadata, performance at scale), leveraging domain-specific managed services to accelerate development, designing stateless, horizontally scalable architectures, and maintaining strong partnerships between healthcare domain experts and cloud infrastructure providers. While the presentation focuses more on infrastructure and clinical workflows than explicit LLMOps practices, the architectural patterns demonstrate production ML deployment considerations: staging data for quality verification, integrating inference into workflow pipelines, enabling model training on consolidated datasets, and supporting real-time predictions with sub-second latency requirements. The mention of HealthScribe indicates generative AI integration, though details are limited. The dramatic workflow improvements (11+ hours to 36 minutes) and clinician satisfaction metrics (100% not wanting to return to microscopes) suggest that when properly implemented with appropriate infrastructure, AI-enhanced diagnostic workflows can achieve both productivity gains and clinical acceptance. The ability to collaborate across geographies and specialties through cloud-based platforms represents a fundamental transformation in how diagnostic medicine operates, with implications for access to specialized expertise and research collaboration. Organizations pursuing similar AI/ML initiatives in healthcare or other highly regulated industries can draw lessons about the value of domain-specific managed services, the importance of addressing data foundations before layering AI capabilities, and the potential for cloud infrastructure to enable entirely new workflow paradigms rather than simply replicating existing processes in a cloud environment. The partnership approach between domain experts and cloud providers appears critical for navigating regulatory complexity while achieving technical innovation.