How to use this list
Topics 1–33 are aimed mainly at medical students, topics 34–56 at computer science students. Many topics work well for a medical and a computer science student together; see joint work.
The topics are a starting point, not a finished assignment. Settle the final wording with your supervisor. Each topic says what kind of work it suits and which data it builds on.
The note “needs ethics approval” means the study involves people: a survey, a teaching experiment or an analysis of the annotators’ own performance. Such work can start only after your university’s ethics committee approves it.
Many topics depend on how far the project has progressed. Topics on open datasets can be taken now; topics on the project’s own annotation open up once pilot annotation begins.
Anatomy and operative surgery
-
“How often is the critical view of safety achieved in laparoscopic cholecystectomy? An analysis of the open Endoscapes2023 dataset”
How many frames before clipping meet each of the three safety criteria, and how often experts disagree.
Term paper, abstract · Endoscapes2023 -
“Calot’s triangle on screen: an atlas of laparoscopic anatomy from annotated frames”
A teaching atlas of how the cystic duct, the cystic artery and the hepatocystic triangle appear.
Thesis, teaching aid · Endoscapes2023, CholecSeg8k -
“Timing a cholecystectomy: the duration and variability of surgical phases”
How long each of the seven phases lasts, which phases most often run long, and what that may relate to.
Term paper · Cholec80 -
“Two operations, one logic? The temporal structure of laparoscopic cholecystectomy and hysterectomy”
Comparing the order and duration of the steps of two different laparoscopic operations.
Term paper, abstract · Cholec80, AutoLaparo -
“What the surgeon sees in robotic rectal resection: how often anatomical structures are visible, and under what conditions”
Which organs come into view, how often, and how much they are covered.
Term paper · DSAD -
“An instrument map of the operation: which instruments are used in each phase of cholecystectomy”
Linking surgical phases to the instruments in view, as a basis for a teaching diagram and for evaluating models.
Term paper, abstract · Cholec80 -
“Smoke, blood and glare: what makes anatomy hard to recognise in laparoscopic frames”
How common visual obstacles affect the accuracy of human and model annotation.
Thesis · DSAD, CholecSeg8k -
“Landmarks of safe cholecystectomy: definitions in clinical guidelines and in the annotation guides of open datasets”
Do the definitions used to annotate data match what surgeons are taught?
Abstract, literature review · Endoscapes2023 and publications
Annotation quality and human factors
-
“Agreement between medical students annotating anatomical structures”
Fleiss’ kappa and the Dice coefficient for different structures: where students agree and where they do not.
Abstract, article · project annotation · needs ethics approval -
“Year, experience, specialty: does annotation accuracy depend on medical training?”
Students of different years, residents and surgeons compared on the same frames.
Article · project annotation · needs ethics approval -
“An annotator’s learning curve: how outline accuracy changes over the first 20 hours”
When accuracy plateaus, and which structures take longest to learn.
Thesis · project calibration tasks · needs ethics approval -
“Trust, but verify: the tendency to accept automatic pre-annotation without edits”
How often annotators accept a model’s suggestion even when it is inaccurate, and how to counter it.
Article · project annotation with pre-annotation · needs ethics approval -
“Where annotators disagree: anatomical reasons for segmentation disputes”
Classifying the disputed frames in the clinical lead’s log: poor visibility, anatomical variant, unclear definition.
Abstract · the project’s log of disputed cases -
“The cost of a frame: annotation effort for surgical video and how to reduce it”
How long a frame and a video take, and what speeds up the work without losing quality.
Term paper · CVAT task statistics -
“Surgical phases through the eyes of a student and a surgeon: agreement on phase boundaries”
How precisely students mark the start and end of phases compared with the dataset authors’ reference.
Abstract · Cholec80 and project annotation
Artificial intelligence in surgery
-
“Where the network goes wrong: a clinical review of errors in surgical phase recognition”
A doctor’s view of a model’s mistakes: which are dangerous and which are harmless.
Thesis · Cholec80, project baseline models -
“Across hospitals: how much phase recognition drops on another hospital’s data”
A model trained on videos from one hospital is tested on videos from others.
Article (with an engineer) · Cholec80, HeiChole -
“Can a model recognise the critical view of safety?”
How closely a model’s assessment matches the experts’, and what explains the differences.
Thesis, article · Endoscapes2023 -
“Does the model understand anatomy? Attention maps and clinical landmarks”
Which parts of the frame the model “looks at”, and whether they match what a surgeon looks at.
Thesis · CholecSeg8k, Endoscapes2023 -
“Pig or human: can surgical vision models learn from operations on animals?”
How models trained on recordings of operations on pigs perform on operations on people.
Article · EndoVis 2017, EndoVis 2018, CholecSeg8k -
“General-purpose segmentation models on surgical video: what they see without fine-tuning”
Testing models such as Segment Anything on laparoscopic frames: what works out of the box and what does not.
Term paper · CholecSeg8k, DSAD -
“An AI assistant in the operating room: which prompts surgeons find useful and which dangerous”
A survey of surgeons and residents on desirable and unacceptable features of intraoperative AI.
Abstract · survey · needs ethics approval
Other specialties
-
“Instruments in cataract surgery: what is hardest for a model to recognise, and why”
An error analysis of instrument segmentation on frames of operations under the microscope.
Term paper, ophthalmology · CaDIS -
“The missed polyp: a clinical review of polyp segmentation errors”
Which polyps models miss most often, and how that relates to their shape, size and location.
Term paper, gastroenterology · Kvasir-SEG, HyperKvasir -
“Through the microscope and through the laparoscope: computer vision tasks in eye surgery and abdominal surgery”
A review of how data, tasks and difficulties differ between the two fields.
Abstract, literature review · CATARACTS, Cholec80
Medical education
-
“An annotated surgical video as a teaching aid for topographic anatomy: design and evaluation”
Building a teaching module on annotated frames and testing its value in class.
Thesis · project annotation · needs ethics approval -
“Case exposure in numbers: does watching annotated operations speed up recognition of anatomical landmarks?”
Comparing the speed and accuracy of recognising structures before and after a series of viewings.
Article · needs ethics approval -
“Annotation as learning: what students learn about anatomy and surgery by annotating operations”
Self-assessment and a knowledge test before and after a semester in the project.
Abstract · survey · needs ethics approval -
“A ‘Find the cystic duct’ trainer: an interactive exercise on open data”
Building a web exercise where a student marks a structure and the program compares the answer with the reference.
Term paper (with a computer science student) · CholecSeg8k, Endoscapes2023
Ethics, law and how science is organised
-
“Can AI learn from YouTube? Ethical and legal questions in using public recordings of operations”
Copyright, patient consent and de-identification: what lawful use requires.
Abstract, review -
“Licences of open surgical datasets: what a researcher is allowed to do”
Comparing the terms of popular datasets and what they mean for publications.
Abstract, review -
“What the frame reveals: personal data in recordings of endoscopic operations and how to remove it”
What information about patients and staff appears in recordings, and how to hide it reliably.
Term paper · project methodology -
“Open science in surgery: how medical students can take part in reproducible research”
The project’s experience as a model of student participation in science.
Abstract, essay
Computer vision and machine learning
-
“A baseline anatomy segmentation model on CholecSeg8k: a reproducible comparison of U-Net, DeepLabV3+ and SegFormer”
The same splits, training budget and metrics for three architectures, with ablations and confidence intervals.
Term paper, thesis · CholecSeg8k -
“Surgical phase recognition in cholecystectomy: temporal convolutional networks versus transformers”
Which architecture handles the long context of an operation better, and where both fail.
Thesis · Cholec80 -
“Little annotation, lots of video: semi-supervised and self-supervised learning for surgical segmentation”
How to use thousands of unlabelled frames to improve a model trained on hundreds of labelled ones.
Thesis, master’s thesis · Endoscapes2023, Cholec80 -
“Adapting Segment Anything 2 to laparoscopic video: fine-tuning and evaluation”
How much fine-tuning a general-purpose model improves organ and instrument segmentation, and how much annotation it takes.
Thesis · CholecSeg8k, Endoscapes2023 -
“Robust to smoke, glare and blur: a corruption suite for testing surgical vision models”
Synthetic and real corruptions, the drop in quality, and ways to reduce it.
Term paper · CholecSeg8k, DSAD -
“When a model should say “I don’t know”: confidence calibration for segmentation models”
Measuring and improving calibration, and abstaining on doubtful frames.
Thesis · CholecSeg8k, Endoscapes2023 -
“From pig to human: domain adaptation for instrument segmentation”
How to transfer a model trained on operations on animals to recordings of operations on people.
Article · EndoVis 2017, EndoVis 2018, CholecSeg8k -
“Under 50 milliseconds: distillation and quantisation for real-time models”
How to keep quality within the target of 30–50 ms per frame, and how to measure latency honestly.
Thesis · CholecSeg8k, Cholec80 -
“Detecting and tracking surgical instruments in video”
An instrument detector and tracker, tracking metrics, errors under occlusion.
Term paper · CholecT50, Cholec80 -
“Recognising instrument–verb–target action triplets”
Models for action triplets and an analysis of the hardest classes.
Thesis · CholecT50
Data, annotation and MLOps
-
“Active learning for annotation: which frames to send to annotators first”
Frame selection strategies and how many annotation hours they save for the same model quality.
Thesis · Cholec80, CholecSeg8k -
“Automatic quality control for annotation: finding suspicious masks before review”
Rules and models that flag likely annotator mistakes.
Term paper · project annotation -
“A library of inter-annotator agreement metrics for segmentation and temporal annotation”
Fleiss’ kappa, Dice, agreement on phase boundaries, with tests and documentation.
Term paper · synthetic data, project annotation -
“A reproducible training pipeline: from a data version to a model card”
DVC, Hydra and MLflow in one pipeline, so any result can be repeated with one command.
Term paper, thesis · any open dataset -
“Leak-free splits: automatically checking that no patient appears in both training and test”
How leaks inflate metrics and how to catch them with tests.
Term paper · Cholec80, Endoscapes2023 -
“An honest benchmark: confidence intervals, repeated runs and protecting the test set”
How to build a leaderboard where results cannot be tuned to the test set.
Thesis · open datasets
De-identification and security
-
“Text on endoscopy equipment screens: detecting and hiding it in a de-identification pipeline”
Finding dates, names and numbers in the frame and masking them reliably.
Term paper, thesis · synthetic frames -
“Out-of-body frames: detecting staff faces and the operating room in recordings”
An inside/outside frame classifier and face detection for automatic filtering.
Term paper · Cholec80 -
“How to prove a video is de-identified: a method for verifying a de-identification pipeline”
Test sets, miss-rate metrics and independent acceptance.
Thesis · synthetic data
Software engineering and interfaces
-
“A pre-annotation model in CVAT as a serverless function: latency, load and scaling”
Integrating a model with the annotation tool and measuring how it copes with many annotators.
Thesis · CVAT, Segment Anything 2 -
“A web anatomy trainer for medical students on open data”
An interactive exercise: the student marks a structure and the program compares the answer with the reference.
Term paper · CholecSeg8k, Endoscapes2023 -
“An interface for clinical review of model errors”
How to show a doctor where and why a model went wrong, so a review takes minutes rather than hours.
Term paper · Cholec80, CholecSeg8k -
“Infrastructure as code for medical AI: data isolation, audit logging and reproducible deployments”
How to build a Kubernetes platform where medical data does not leak and every deployment is repeatable.
Thesis · project infrastructure, by arrangement
Joint work for medical and computer science students
The most interesting work happens in pairs: a medical student owns the clinical meaning, a computer science student owns the model and the code. Topics that pair well:
- 16 + 35: a phase recognition model and a clinical review of its errors;
- 20 + 40: transferring models from operations on animals to operations on people, from the engineering and the clinical side;
- 29 + 54: the “Find the cystic duct” trainer, its content and its implementation;
- 13 + 45: disputed frames through a surgeon’s eyes and automatic detection of annotation errors;
- 16 + 55: an interface in which a doctor reviews a model’s errors.
Rules for work on project data
- Cite the dataset authors and follow the licence terms. Most datasets allow non-commercial use only. See the licences page.
- No patient data beyond the published datasets and permitted videos. Do not use recordings from your own hospital.
- Studies involving people only after approval from your university’s ethics committee and with the participants’ consent.
- A negative result is a result. Work that honestly shows an idea did not work is valuable to the project.
- Publications using project data must be agreed with the project’s scientific lead in advance. Authorship follows the contribution agreement.
Suggest your own topic
Have an idea that is not on the list? Write to support@opensurgicalvision.org with the subject “Research topic” or start a thread in GitHub Discussions. Good topics will be added to the list.
To get started, see the pages for medical students, for computer science students and for educators.