Senior ML engineer on Midjourney's clinical team, owning tissue-class segmentation and labeling models for an ultrasound CT clinical analysis layer — building training/eval pipelines and datasets with clinicians, and productionizing a versioned, HIPAA-bound inference service. Core stack: PyTorch, segmentation architectures (U-Net/nnU-Net, SAM), and tools like MONAI and ITK.
Own the tissue-class segmentation and labeling models for the ultrasound CT clinical analysis layer, and the pipelines that make them trainable and verifiable.
Retune across 2D per-slice, 3D volumetric, and 2D×3D fusion as reconstructed image inputs are continuously updated, and clinical indications for use expand.
Define training/evaluation pipelines, datasets, and metrics from the ground up or from open source; map model behavior to user needs and design requirements.
Work with data labeling contractors, expert clinicians, and our internal cloud/data teams on labeling specs, QC, and dataset versioning.
Help productionize models into a versioned, HIPAA-bound analysis service: reproducible/low-latency inference, per-prediction confidence, drift monitoring, and safe fallbacks.
What we’re looking for
Strong applied ML experience with a track record of developing new models — architecting, training, and evaluating from scratch as well as benchmarking against existing models.
Experience with image segmentation (semantic/instance, 2D and ideally 3D/volumetric) and the modeling and training-data choices that make it robust across diverse patient anatomy.
Comfortable moving fluidly between open-ended research iteration and producing quantifiable, testable models.
Fluent in modern deep-learning tooling (e.g., PyTorch) and current development practices.
Comfortable working under design controls, where model changes carry documentation and verification weight.
Useful experience
Image segmentation and label generation with modern architectures (U-Net / nnU-Net, 3D U-Net, transformer-based and promptable segmentation like SAM), including the geometry that ties voxel- and mesh-level predictions back to a coordinate frame.
Learning under limited or noisy supervision: self-supervised / semi-supervised methods (masked autoencoders, contrastive pretraining like DINO/SimCLR), active learning, weak labels, and simulation-driven pretraining.
Hands-on experience with data curation for ML: building datasets from messy, real-world sources, helping to define ground truth, and managing labeling or simulation pipelines (MONAI, ITK / SimpleITK, 3D Slicer).
Experience with segmentation models for ultrasound imaging, whether on synthetic or real images
ML for imaging or inverse problems in physics-based domains (CT, MRI, ultrasound, or adjacent), and comfort working alongside reconstruction/signal-processing teams.
Deploying models in versioned, auditable, high-stakes settings.
A background in anatomy, medical imaging, or body composition and prior work with existing segmentation models is a plus.