INTL
Freelancer
전문가
외주
원격 가능
AWS Deployment of Dual-Model
예산
$30~$250 AUD
예상 기간
1~2개월
난이도
전문가
기술 스택
Python
Machine Learning (ML)
Deep Learning
Model Deployment
Amazon Web Services (AWS)
AWS SageMaker
AWS Lambda
AWS API Gateway
AWS S3
AWS Route 53
AWS ACM
IAM
Hugging Face
DICOM
3D Volumetric Reconstruction
API Development
JSON
AI 분석 요약
이 프로젝트는 AWS 클라우드에 의료 영상 AI 듀얼 모델(Pillar-0 및 Sybil-1.5)을 배포하는 파이프라인을 구축하는 것입니다. DICOM 파일을 처리하여 질병 분류 및 폐암 위험 예측을 수행하고, 안전한 API 엔드포인트를 통해 결과를 제공합니다. AWS SageMaker, Lambda, API Gateway 등의 클라우드 서비스 활용 능력과 딥러닝 모델 배포 및 의료 데이터 처리 지식이 요구됩니다.
프로젝트 원문 설명
deploy a dual-model pipeline on AWS.
Scope of Work
Dual-Model Deployment:
Deploy the Pillar-0 (Atlas architecture) for multi-finding classification across Chest, Abdomen, and Brain.
Integrate Sybil-1.5 for specialized future lung cancer risk prediction (1–6 year horizon).
Inference Pipeline & Report Generation:
deploy pipeline that takes zipped DICOM files, performs 3D volumetric reconstruction, and runs concurrent inference. inference.py script (attached). This script shows how to read the (Zip file), stack of the medical pictures (DICOMs) into a 3D block, and run both the Pillar-0 model (for findings) and Sybil-1.5 (for cancer risk)
requirements.txt -This tells AWS which specialized tools to install to read 3D medical images
AWS Cloud Architecture:
Host models on AWS SageMaker (GPU instances like ml.g5.2xlarge).
Implement an API Gateway + Lambda front-door.
Support the "Locker Pattern" for direct S3 uploads of large volumes (>10MB) via pre-signed URLs. Model Resources
Pillar-0 Collection: huggingface.co/collections/YalaLab/pillar-0
Sybil-1.5 Model: huggingface.co/YalaLab/Pillar0-Sybil-1.5 Deliverables
A live, secured AWS API endpoint.
A sample "Radiology Style" JSON and PDF output generated from a test CT.
When you upload to AWS, your bucket MUST look like this for the code:
s3://your-s3-bucket/
└── models/
└── dxpert-v1/
└── model.tar.gz <-- THIS CONTAINS EVERYTHING BELOW:
├── code/
│ ├── inference.py
│ ├── atlas.py (The Model architecture code)
│ └── requirements.txt
├── pillar0_weights.pth
└── sybil15_weights.pth
Atlas Architecture: pull the Atlas class from the YalaLab GitHub/HuggingFace repo and include it in the code/ folder.
Binary Media Types: enable application/zip in the API Gateway Settings,
GPU Instance: deploy on a g5.2xlarge
S3 Permissions: Ensure the IAM Role assigned to your SageMaker endpoint has the AmazonS3ReadOnlyAccess policy so it can reach into the bucket and grab the zip files.
The "Locker Pattern" for Large Volumes (500MB):
Because our CT volumes exceed AWS API Gateway limits (10MB), you must implement a Pre-signed S3 URL workflow.
Develop a Lambda-based "Key Maker" that generates temporary upload links for clients to push 500MB zip files directly to S3.
Configure a Custom Domain Name (api.ct-expert.ai) for the API.
Manage the full networking stack: AWS ACM (SSL/TLS), API Gateway, and Route 53 (Alias records).
Ensure the endpoint is secure and production-ready
Scope of Work
Dual-Model Deployment:
Deploy the Pillar-0 (Atlas architecture) for multi-finding classification across Chest, Abdomen, and Brain.
Integrate Sybil-1.5 for specialized future lung cancer risk prediction (1–6 year horizon).
Inference Pipeline & Report Generation:
deploy pipeline that takes zipped DICOM files, performs 3D volumetric reconstruction, and runs concurrent inference. inference.py script (attached). This script shows how to read the (Zip file), stack of the medical pictures (DICOMs) into a 3D block, and run both the Pillar-0 model (for findings) and Sybil-1.5 (for cancer risk)
requirements.txt -This tells AWS which specialized tools to install to read 3D medical images
AWS Cloud Architecture:
Host models on AWS SageMaker (GPU instances like ml.g5.2xlarge).
Implement an API Gateway + Lambda front-door.
Support the "Locker Pattern" for direct S3 uploads of large volumes (>10MB) via pre-signed URLs. Model Resources
Pillar-0 Collection: huggingface.co/collections/YalaLab/pillar-0
Sybil-1.5 Model: huggingface.co/YalaLab/Pillar0-Sybil-1.5 Deliverables
A live, secured AWS API endpoint.
A sample "Radiology Style" JSON and PDF output generated from a test CT.
When you upload to AWS, your bucket MUST look like this for the code:
s3://your-s3-bucket/
└── models/
└── dxpert-v1/
└── model.tar.gz <-- THIS CONTAINS EVERYTHING BELOW:
├── code/
│ ├── inference.py
│ ├── atlas.py (The Model architecture code)
│ └── requirements.txt
├── pillar0_weights.pth
└── sybil15_weights.pth
Atlas Architecture: pull the Atlas class from the YalaLab GitHub/HuggingFace repo and include it in the code/ folder.
Binary Media Types: enable application/zip in the API Gateway Settings,
GPU Instance: deploy on a g5.2xlarge
S3 Permissions: Ensure the IAM Role assigned to your SageMaker endpoint has the AmazonS3ReadOnlyAccess policy so it can reach into the bucket and grab the zip files.
The "Locker Pattern" for Large Volumes (500MB):
Because our CT volumes exceed AWS API Gateway limits (10MB), you must implement a Pre-signed S3 URL workflow.
Develop a Lambda-based "Key Maker" that generates temporary upload links for clients to push 500MB zip files directly to S3.
Configure a Custom Domain Name (api.ct-expert.ai) for the API.
Manage the full networking stack: AWS ACM (SSL/TLS), API Gateway, and Route 53 (Alias records).
Ensure the endpoint is secure and production-ready
Freelancer에서 원본 확인
원본 보기