CIMILab-Deploy

Problem

Translating a well-performing research model into a reliable clinical tool requires engineering that goes far beyond model training: efficient inference, serving infrastructure, monitoring for distribution shift, and optimization for low-resource settings. CIMILab-Deploy bridges this gap.

Approach

CIMILab-Deploy will provide deployment-ready wrappers and tooling for CIMILab's trained models:

• Optimized inference: ONNX export, TensorRT optimization, quantization pipelines (INT8/FP16)
• Serving infrastructure: FastAPI-based REST endpoints with async batch inference
• Uncertainty-aware outputs: Every deployed model exposes confidence estimates alongside predictions
• Edge adaptation: Lightweight model variants for deployment on resource-constrained hardware in low-resource clinical settings
• Monitoring hooks: Data drift detection, performance degradation alerts

Status

CIMILab-Deploy will initially wrap the DSANet-ISLES segmentation model. The existing deploy_DSANet_ISLES22 repository provides the foundation; this project generalizes it into a reusable framework.