As an intern, you join the core DL team and contribute at one or both layers of our training stack: Pretraining — learn generalizable representations on large, de-identified datasets; Fine-tuning — adapt those representations (and VLMs) to clinical tasks and ship improvements. Work is prioritized by expected product and clinical impact; research is a means to that end.

Where you’ll contribute

Pretraining (images ± text) Design and scale representation learning for 2D/3D medical imaging:

• Objectives: masked image modeling, self-distillation/contrastive (MAE/DINO-style), vision–language alignment (CLIP-style with radiology reports).
• Modalities/architectures: X-ray, CT, occasional MRI; 2D/3D ViTs and UNet-style decoders.
• Systems: high-throughput DICOM loaders, strong augmentations, mixed-precision, distributed training.
• Evaluation: transfer to target tasks, label-efficiency curves, robustness across sites/vendors.

Fine-tuning (product models & VLMs) Adapt and optimize models that power our products and workflows:

• Core tasks: detection/segmentation/registration; follow-up (temporal matching, lesion tracking, measurements); calibration & uncertainty.
• VLMs: image encoder → decoder LLM for report generation/summarization/structured extraction. Techniques include supervised fine-tuning on paired image–report data, instruction tuning, alignment of visual tokens (e.g., resamplers/Q-Former-style adapters), and LoRA/PEFT on both vision and language components.
• Efficiency & deployment: distillation, pruning/quantization, KV-cache and batching for throughput; ONNX/TensorRT inference.
• Evaluation: AUROC/FROC/Dice, ECE calibration; for VLMs—finding-level label agreement, RadGraph-style entity/relational metrics, factuality checks vs imaging labels; clinician review.

You’ll spend time where it moves metrics most. Some interns focus on pretraining, others on fine-tuning/VLMs; many touch both.

How we work (engineering standards)

• Reproducibility: Hydra configs, seeded runs, model registry; tracked experiments.
• Production-ready code: typed Python, tests on data/metrics, documented PRs, code review.
• Measured progress: clear win criteria on accuracy, generalization, latency, and memory.

Our stack

PyTorch (Lightning), MONAI, timm/Hugging Face; NumPy/scikit-image; DICOM tooling; Weights & Biases (W&B) / ClearML/DVC; multi-GPU training; ONNX/TensorRT for inference; containerized services.