MONAI-based imaging pipelines Imaging
Building reproducible inference pipelines on Project MONAI for chest X-ray, CT, and MRI cohorts — segmentation, classification, and anomaly detection as upstream signals for downstream report generation.
AI-assisted second opinions for medical imaging.
An early-stage initiative exploring how Project MONAI, open-source large language models, and cloud-native infrastructure can deliver structured second-opinion radiology reports to assist clinicians.
Focus Areas
What this initiative explores.
Six intersecting questions at the boundary of medical imaging AI, large language models, and clinical decision support.
LLM-driven report synthesis NLP
Using open-source LLMs to convert structured imaging findings into clinician-readable second-opinion reports, with explicit grounding and uncertainty quantification.
Multimodal grounding & provenance Safety
Linking each generated finding back to the source slice, region of interest, and supporting literature — making AI-generated radiology reports auditable rather than opaque.
Cloud-native deployment Infrastructure
Architecture patterns for running MONAI plus LLM workloads cost-effectively on major hyperscaler platforms — balancing GPU spend, latency, and storage for DICOM-scale data.
Calibration & clinical concordance Evaluation
Benchmarking generated second-opinion reports against board-certified radiologist readings — measuring agreement, false-negative rate, and clinically actionable disagreement.
Open-source ecosystem & reproducibility Open
Favouring open weights, open datasets, and reproducible inference recipes wherever possible — so findings can be independently verified rather than trusted on faith.
Technical Stack
Built on open and trusted foundations.
Three layers, each chosen because the medical imaging community has already validated it.
Project MONAI
The PyTorch-based, NVIDIA-backed open framework for medical imaging AI. Domain-specific transforms, network architectures, and training loops purpose-built for DICOM and NIfTI data.
Open-Source LLMs
Llama, Mistral, and medical-domain fine-tunes. Used for report drafting, finding summarization, and clinical-language alignment with explicit grounding.
Cloud-Native Infra
GPU-backed inference on hyperscaler infrastructure, object storage for DICOM cohorts, and event-driven workflows for batch and on-demand second-opinion generation.
How It Works
From pixels to clinician-ready report.
A four-step pipeline, designed for auditability at every stage.
Ingest
DICOM or NIfTI study uploaded to object storage. Metadata extraction, anonymization, and integrity check.
Analyze
MONAI inference pipeline runs segmentation and classification. Findings emit as structured JSON with confidence scores.
Synthesize
Open-source LLM converts structured findings into a draft second-opinion report, grounded with citations to source slices.
Review
Output presented for clinician review — never as autonomous diagnosis. All decisions remain with the licensed physician.
Get in touch
For collaboration, dataset access discussions, or to share related work.
info@2nd-opinion.click