research
What a model can carry from one setting to another
Collaborative surgical-AI research on generalizability, annotation, and the limits of a promising result.
- Role
- Surgical AI Research Lead · co-author
- Status
- Ongoing research · public conference abstracts
Context and real problem
Pulmonary-vasculature segmentation is a research problem shaped by changing anatomy, limited annotated video, and the difficulty of transferring a method between settings. A result on familiar data does not settle how a model will behave elsewhere.
Andres’s role
I have contributed to surgical-video annotation, evidence synthesis, evaluation work, and collaborative research writing. My continuing Surgical AI Research Lead role moved with the lab from Tufts to UPMC; the earlier assistant role and the Harvard/MGH research collaboration remain distinct parts of the chronology.
Collaborators and setting
The public conference abstracts list multidisciplinary teams associated with Tufts Medical Center and Plaksha University. My earlier video-annotation work was part of a Harvard Medical School / Massachusetts General Hospital research collaboration. These are shared contributions, with clinical and technical expertise coming from different collaborators.
Methods and tools
The public abstracts describe annotated surgical video, CVAT, segmentation-model evaluation, and tests across anatomical settings. My annotation work has used MOSAIC and CVAT. The team’s methods include Mask R-CNN; separate research versions also use XMem. These methods belong to distinct study records and should not be collapsed into one benchmark.
Responsibility and safety
Annotations require clinical interpretation, and evaluation must account for changes in setting. Surgical footage and private datasets stay outside this website. A segmentation score is evidence about a defined task; it does not demonstrate safer operations or clinical readiness.
Outputs and outcomes
I co-authored the 2025 ISMICS and STSA conference abstracts and a related proceedings abstract in The Van Wickle Journal for the National Undergraduate Research Conference at Brown. These collaborative outputs examine generalizability and annotation demands; they establish a research contribution, rather than clinical readiness.
What I learned
The question of generalizability changes how I read a result. I want the conditions of a comparison, the annotation decisions, and the limits of transfer to remain visible alongside the model’s performance.
Current status and links
The conference abstracts are public. I continue in the lead role following the lab’s move to UPMC in September 2026. The links provide the shared research record; surgical footage, source datasets, and unpublished findings remain within their approved setting.