Adherence to and compliance with local laws and regulations in data capture, storage, and use are just as important as compliance with accepted ethical guidelines around research.
True anonymization and full deidentification of data are difficult to achieve. Appropriate measures must be in place to ensure optimal protection of patient privacy.
To obtain a reliable result in your research, establishing a well-planned database structure for storage and access to data is as important as your machine learning model and analysis.
Computer vision is a young field. Its current era started in 2012 with Krizhevsky’s application of deep convolutional neural networks,1 and this “deep learning revolution” made accurate image and video classification a reality with a resultant explosion in computer vision research (nearly 30,000 publications in 2017). The medical community, though, has underperformed in contributions, with less than 1000 annual publications in artificial intelligence.2 Computer vision medical research, particularly that focused on surgery, is therefore an even smaller slice of the pie.
With a paucity of research groups, there is no definitive roadmap to creating an effective medical computer vision group. New labs will inevitably encounter hurdles and unforeseen complications. Our group, the Surgical Artificial Intelligence and Innovation Laboratory (SAIIL) at the Massachusetts General Hospital (MGH), is one of just a handful of groups that focus on computer vision and surgery. Using our laboratory’s experience, this chapter will provide a roadmap to navigating these poorly charted waters.
Prior to embarking on a computer vision study in surgery, a researcher must first determine a question to address. This question will be unique to each group’s interests, with common computer vision problems including operative temporal segmentation, phase recognition, tool recognition, operative skill assessment, and clinical decision support. Once there is a question, one must then identify the appropriate technical expertise in answering the question.
Chapters 6, 7, 8, and 10 in particular touch on various research questions that have been tackled in the field of surgical computer vision.
As with any research, the selection of appropriate methodology is critical. Thus, the first recommendation for anyone looking to engage in computer vision research is to partner with an expert in, at least, the technical aspects of the field. Guidance from a PhD-trained expert will help you refine your question to be specific, measurable, and achievable while also preventing you from falling into common traps regarding data set selection and model evaluation.
Although the specifics for any particular research project will vary, the remaining chapter will review the 3 major considerations when conducting research in surgical computer vision:
Data acquisition and preparation
DATA ACQUISITION AND PREPARATION
Machine learning and computer vision models live or ...