Use case Surgomics (Computer-assisted surgery)

NCT, DELL, TUD & SCONTAIN

EXTREME NEAR-DATA PROCESSING PLATFORM

Description of the use-case

Surgomics/computer-assisted surgery aims to improve patient outcome by applying machine learning along the entire treatment path

Surgomics

Description of the use-case

By analysing data along the surgical path, it becomes possible to assist the surgical staff, e.g. via highlighting relevant structures or via robot assistant but also to predict the possible outcomes.

Surgomics

Why extreme data?

Data processing in computer-assisted surgery needs to be:

• Real-time
• Low-latency
• Secure

Surgomics

Important for patient safety

Important for data protection

Characteristics data/streams

• Large (HD/4K/Stereo video)
• Variable frame rates
• 25-60 FPS for videos
• 100-1000 FPS for other sensors, e.g. robot
• Highly variant
• Contains personal health information
• Synchronization of streams and results

Components integrated

Surgomics

Datasets

Surgomics

Data spaces

Surgomics

Intraoperative Data

Storage

Annotation

Model Training

Online Processing

Surgeon's Display/AR/Robot Control….

Demo Federated Learning

Surgomics

• Problem:
• High variance data at multiple centers
• Data protection restricts sharing
• Goal: Secure and confidential ML model training without sharing raw data between centres
• Solution: Scone (Scontain’s secure computing solution), Flower (Federated Learning framework), Docker, Pytorch, Python

Demo Federated Learning

Surgomics

Problem: Confidentiality, integrity and consistency if adversary has root and hardware access?​

Approach: Run applications in Trusted Execution Environments (TEE) without access by root users

​

Demo Federated Learning

Surgomics

Problem:

• A large number of malicious clients may collude to reveal local data and models of the other clients
• Malicious clients can tamper with local data or parameter updates to corrupt the global mode

Approach: Develop a Data Broker component based on the​ Remote attestation mechanism of TEE​

​

​

Demo Federated Learning

Surgomics

Executing of the federated learning demo inside the SGX enclaves using SCONE’s network shield, configured through the Data Broker’s CAS(Configuration and Attestation Service) component.

Confidential Computing Framework

With the new infrastructure provided by KIO for the next phase, it would be possible to integrate Flower into the Confidential Computing Framework that is being prepared

Next Phase of the Project

KPIs FL

KPI-4 - High levels of data security and confidential computing validated using TEEs and federated learning in adversarial security experiments.

Surgomics

Demo Streaming /w Pravega

Surgomics

• Problem:
• Low-latency data processing
• Scalability to multiple ORs
• Make live data accessible for training
• Goal: Platform for live streaming and real-time processing for surgical data, while storing videos for later use. Adaptability to changing demands
• Solution: Containerize components, integrate into Pravega cluster, determine needs via learning from past usage

Streaming Infrastructure PoC

Surgomics

Demo Streaming /w Pravega

Surgomics

KPIs Streaming

• KPI-2 - Latency and data access improvements:
• Containerized AI applications using Pravega can achieve end-to-end IO latency 6-9ms at p95.
• This IO end-to-end latency is up to 45% lower than inference latency (p95).
• Therefore, our PoC meets the latency requirements

​

Surgomics

KPIs Streaming

• KPI-3 – Resource auto-scaling:
• 6-hour fluctuating workload experiment.
• Scale stream segments at 0.8K events/second (stream policy).
• Keep # stream segments = # Flink task managers (orchestrator).
• Simple approach to keep Pravega stream IO and Flink compute capacity aligned automatically, which simplifies management.

Surgomics

KPIs Streaming

• KPI-5 - Simplicity and productivity:
• Containerization of AI inference in our PoC is estimated to reduce by ~60% the deployment and data management time.
• Our PoC can help increase productivity by simplifying deployment and data management.

​

Surgomics

Thank you