St. Jude Opens a Postdoctoral Role Focused on AI for Medical Imaging
St. Jude Children’s Research Hospital is recruiting a postdoctoral research associate to build machine learning, deep learning, and foundational AI systems for medical imaging, with radiology at the center of the work and biomedical informatics running alongside it. The position is based in Memphis, Tennessee, and is framed as a research-heavy appointment for candidates who want to turn fast-moving AI methods into tools that can actually be used on imaging problems. The lab led by Dr. Pritam Mukherjee is looking for a postdoc who follows the current AI literature closely and can translate advances in large language models, multimodal foundation models, and vision-language systems into practical imaging research.
This is not a narrow coding post. The remit extends from segmentation and detection to quantification, volumetrics, density measurements, and image-based biomarker discovery across CT, MRI, and X-ray. In a field where model novelty often outruns clinical usefulness, the job description makes the expected direction explicit: build systems that address real imaging questions and stay grounded in what radiologists, clinicians, and engineering teams can actually use.
Multimodal AI Meets Quantitative Imaging Workflow
The role sits at the intersection of modern computational methods and conventional radiology workflow. Applicants are expected to design and implement 2D and 3D architectures, including CNNs, transformer-based models, and vision-language or foundational models. The lab also wants someone who can apply and adapt large language models to biomedical and imaging-adjacent tasks, including training, fine-tuning, inference, and deployment. In practical terms, this means the postdoc may be working across both image-centric pipelines and text-heavy biomedical problems, an increasingly common pairing in translational AI.
Quantitative imaging sits at the center of that agenda. Measures such as tumor burden, lesion volume, tissue density, or other extracted biomarkers can make imaging more reproducible and more useful for longitudinal research, especially when datasets span multiple institutions. Those outputs matter because they can shift imaging from descriptive interpretation toward structured phenotyping, where models help standardize what is measured rather than merely labeling images. In pediatric research settings, where precision and consistency are especially valuable, that type of pipeline can shape downstream study design as much as it influences analysis.
Large Datasets, Curated Inputs, and the Burden of Reproducibility
The lab is described as data-rich and well resourced, with access to large multi-institutional datasets and St. Jude’s high-performance computing infrastructure. Some datasets will already be curated, while others will require annotation and internal curation work. That combination is clinically and scientifically important because model quality in medical imaging still depends heavily on label fidelity, dataset diversity, and the handling of site-specific variation in scanners, protocols, and acquisition quality.
A lot of AI work fails quietly at this stage. A model can look strong on a single curated dataset and then struggle when the image distribution changes, the annotation standard shifts, or the deployment context becomes more heterogeneous. The job description repeatedly points toward reproducibility and traceability: versioning data, models, and code; building scalable preprocessing and training pipelines; and using collaborative development practices such as Git, testing frameworks, and code review. Those details are not administrative extras. They are the difference between a publishable prototype and a research platform another team can extend.
Research Culture Built Around Independence and Mentorship
The position also carries an unusually strong academic development component for a postdoctoral role. Support for independent funding is built into the structure, including encouragement for applicants interested in a K99/R00 or a similar career development award. For a postdoc in computational medicine, that matters because funding trajectory often determines whether a trainee remains in academia long enough to build a durable research program.
There is also room for mentorship experience, although it is optional rather than mandatory. Candidates who want to supervise or support junior lab members may gain experience with PhD students or research assistants. That arrangement fits a lab environment where technical depth and scholarly independence are both expected. The advertised appointment runs two to three years, with a flexible start date, giving enough runway for a serious body of work without locking the candidate into a rigid timeline.
What the Position Requires in Practice
The required background is broad but specific. St. Jude is seeking a PhD or MD, or MD/PhD, in fields such as computer science, data science, biomedical engineering, medicine, or a related discipline. The core technical requirements include demonstrated work in segmentation, detection, or quantitative imaging algorithms, strong Python skills, and fluency with deep learning frameworks such as PyTorch or TensorFlow. Experience with modern architectures, including U-Net variants, detection frameworks, transformers, or foundational models, is part of the expected baseline.
Evaluation literacy is another central expectation. The posting calls for familiarity with metrics such as Dice, IoU, ROC/AUC, sensitivity, and specificity. In medical imaging, those metrics are not interchangeable. Dice and IoU speak to overlap and contour quality, while ROC/AUC and sensitivity or specificity reflect classification performance. A strong postdoc in this space has to understand not only how to optimize them, but when one metric is more honest than another for a given clinical task.
Why Radiology-Centered AI Still Needs Clinical Discipline
The preferred qualifications make clear where the lab sees room for added value. Experience with LLMs, DICOM workflows, microscopy or pathology, biomedical informatics, cloud environments, and translational or regulatory-aware research would all strengthen an application. Clinical exposure is also valued, especially where it helps separate interesting model behavior from outputs that would actually matter at the bedside. That kind of judgment is crucial. A technically elegant system can still miss the clinical target if it fails to produce interpretable, reproducible, or operationally feasible results.
The role is therefore as much about scientific restraint as technical ambition. Radiology AI has a long history of promising more than it can safely deliver when training data are narrow or the downstream use case is vague. Here, the emphasis on clinician collaboration, annotation quality, deployment-aware thinking, and rigorous experiment tracking suggests a more disciplined approach. The job is built for someone who wants to work at the boundary of method development and biomedical application, not for someone looking to optimize benchmarks in isolation.
The advertisement closes with a straightforward institutional position on hiring: technology may assist parts of the process, but recruiting decisions rely on human review and engagement. For applicants, the practical question is whether they can combine model-building speed with research maturity, especially in a setting where multi-institutional imaging data, biomedical informatics, and pediatric translational science intersect. The opening is broad, but the expectations are specific, and the work will likely be judged by whether the resulting systems are reproducible, clinically sensible, and strong enough to support manuscripts, collaborative projects, and independent funding applications.