In short
- OpenSurgicalVision is an open research and education project in surgical computer vision.
- We collect open data, annotate it together with surgeons, train models and publish the results honestly – including the ones that did not work.
- We do not build a medical device and do not give clinical advice.
What follows are our plans, not a list of finished features. The roadmap shows what is actually done.
What we do
Bring open data together. We convert existing open surgical datasets into one format with one way to load them. Where a license does not allow redistribution, we publish download and conversion scripts instead of the data.
Build a reproducible benchmark. The same data splits (by patient, not by frame), the same metrics and an open results table, so that models can be compared fairly.
Train and publish baseline models. Instrument and anatomy segmentation, surgical phase recognition and surgical smoke removal. Every model comes with a description: what it was trained on, where it performs poorly, and how to reproduce it.
Annotate data with surgeons. Practising volunteer surgeons annotate video in CVAT. We measure quality by how well experts agree with each other, and the project’s clinical lead decides disputed cases.
Share our tools. Annotation, training and model-acceleration pipelines and a demonstration service are all open, so others can repeat and improve our work.
Experiment in the open. We change and compare model architectures on open data. Every experiment is described, reproducible and compared with a baseline – negative results included.
Automate routine work. Software assistants prepare data, start experiments and write reports; people keep the decisions.
Write papers and give talks. We turn the results into scientific publications and conference presentations.
What we do not do
We do not build a medical device. The project is not seeking and does not plan to seek registration as a medical device. All models and materials are for research only: RESEARCH USE ONLY – NOT FOR CLINICAL USE.
We do not accept data that can identify a patient. Only de-identified data with confirmed permission for use enters the project. Raw patient video and images are never handed over to us.
We do not compete with medical equipment manufacturers. We work at the level of science and education: simulators, resident training and reproducible research.
We do not use data or work from commercial projects. Contributors’ work for their employers and their clients’ data stay out of the project.
We do not run full automatic architecture search. It needs computing resources an open project does not have. We try variants within a space a person has defined in advance and within a fixed compute budget.
We do not present technical progress as clinical progress. A better metric is an engineering result. The project does not claim that a model is “clinically better”.
Why this matters
- Open data is not free data. Being able to download a dataset does not grant the right to redistribute it. We check the license of every source.
- A “research use only” notice does not replace legal assessment. Legal questions are handled separately, with lawyers.
- A good metric does not prove benefit to patients. Only clinical studies can establish clinical value – a benchmark cannot.