Edge PACS

The foundational property of an Edge PACS is its ability to ingest, store, and serve DICOM objects entirely on local storage without requiring a WAN connection. This ensures that imaging workflows remain operational during network outages. The system maintains a local DICOM database that can be queried via standard C-FIND and C-MOVE services, acting as a fully autonomous mini-PACS. This architecture is critical for point-of-care ultrasound and ambulance-based CT scanners where cloud connectivity is intermittent.

Unlike a traditional PACS that passively stores images, an Edge PACS embeds a hardware-accelerated inference engine directly into the ingestion pathway. As raw DICOM slices are reconstructed, they are immediately passed to a deployed model for tasks like hemorrhage detection or pneumothorax classification. The resulting AI findings—often stored as DICOM Structured Reports (SR) or Presentation States (GSPS)—are written back to the local database and visually overlaid on the study before it is opened by a clinician, achieving zero-click workflow integration.

An Edge PACS must manage multiple diagnostic models within strict thermal and memory constraints. It employs a model server that loads and unloads task-specific networks on demand. Key techniques include:

• Mixed Precision Inference: Running models in INT8 or FP16 on the NPU to maximize throughput.
• Memory Footprint Budgeting: Ensuring the total RAM consumed by the OS, PACS database, and loaded models never exceeds the device's physical limit.
• Concurrent Model Execution: Pipelining organ segmentation and lesion detection models to run in parallel on heterogeneous compute units (GPU + NPU).

For seamless integration with upstream hospital systems, the Edge PACS exposes a DICOMweb API (STOW-RS, QIDO-RS, WADO-RS). This allows a central Vendor Neutral Archive (VNA) or enterprise PACS to query and retrieve studies—including AI-generated secondary captures and reports—using standard HTTP protocols. The Edge PACS acts as a temporary authoritative source, automatically forwarding completed studies to the cloud or central archive when connectivity is restored, ensuring data consistency without manual intervention.

Diagnostic AI models require continuous improvement. An Edge PACS incorporates a secure Over-the-Air (OTA) update client that manages the full lifecycle of deployed algorithms. This includes:

• Atomic Model Swaps: Downloading a new model version to a staging slot and switching over with zero downtime.
• A/B Testing: Running a shadow model alongside the active model to compare performance on live data.
• Rollback Triggers: Automatically reverting to the previous model version if model drift detection identifies a statistically significant degradation in output quality.

A core architectural property is the integration of uncertainty quantification directly into the radiology worklist. The Edge PACS does not just append AI findings; it calculates a confidence score for each prediction. Studies with high-confidence positive findings are flagged as STAT, while cases where the model reports high epistemic uncertainty are automatically routed to the top of the reading queue with a flag for mandatory human review. This creates a safety-critical feedback loop that optimizes radiologist workflow based on model confidence.

The practice of embedding a deep learning inference pipeline directly within or adjacent to a medical imaging modality (MRI, CT, ultrasound). This eliminates network transfer latency by processing raw sensor data immediately.

• Enables real-time image reconstruction and quality control
• Flags critical findings before the patient leaves the scanner
• Requires tight integration with OEM firmware and DICOM acquisition protocols

A compression technique that reduces the numerical precision of a neural network's weights and activations from 32-bit floating point (FP32) to 8-bit integers (INT8) or lower. This dramatically shrinks memory footprint and accelerates inference on resource-constrained edge hardware.

• Post-Training Quantization (PTQ) uses a small calibration dataset
• Quantization-Aware Training (QAT) simulates quantization during fine-tuning for higher accuracy
• Essential for deploying large diagnostic models on embedded PACS servers

A hybrid architecture that intelligently distributes AI workloads between local edge devices and centralized cloud resources. Simple, latency-sensitive inferences execute locally, while complex, computationally intensive analyses are offloaded.

• Triage models run on-device for immediate flagging
• Differential diagnosis or longitudinal comparison models run in the cloud
• Requires robust sync protocols and offline fallback capabilities

Statistical methods that equip a diagnostic AI model with the ability to estimate the confidence of its own predictions. This enables an edge PACS system to automatically flag ambiguous cases for mandatory human review.

• Monte Carlo Dropout: applies dropout at inference time to generate prediction distributions
• Deep Ensembles: uses multiple independently trained models to measure disagreement
• Critical for patient safety and regulatory compliance in autonomous triage workflows

Over-the-Air update infrastructure for wirelessly deploying new or revised AI models, firmware, and software to deployed medical devices. This ensures edge PACS nodes remain current without disrupting clinical workflows.

• Requires A/B partitioning for atomic, rollback-safe updates
• Must comply with FDA change management protocols for SaMD
• Includes cryptographic signature verification to prevent tampering