When I heard SurgeonAI’s vision for the first time I was thrilled and excited. Here I found the opportunity to contribute my AI and computer vision knowledge and experience for saving lives and improving surgical operations results. During the development of several AI algorithms in the laparoscopy domain I found out the necessity of a close collaboration between the expert surgeons and the AI developers. In this blog I would like to share my insights and observations.
The process of the development of AI algorithm can be divided to several steps in which the collaboration of expert surgeons and the AI developers is required. The first step is defining the problem. This task is mainly done by the expert surgeon but it has to be clear and understandable to the AI developer. Therefore, the AI developer should get familiar with the surgical operation field, mainly laparoscopy domain, including the procedures, surgical and medical language, anatomy, surgical tools and the challenging task of reviewing hours of laparoscopy surgeries videos (not recommended before lunch). On the other hand, the AI developer should share with the expert surgeon the technical aspects, such as required input data characteristics, AI algorithms capabilities and limitations, etc.
The next step is to define the requirements (required results and the best way to presents them). Establish mutual language and understanding between the surgeon and AI developer. The main tool for achieving this is using bidirectional knowledge transfer (Surgeon to AI developer and vice versa). The success of this step is crucial and requires full collaboration between the surgeons and the AI developers, bridging the gap between technical jargon and medical terminology which is essential for effective communication.
In the dataset creation step, which is in our case mostly videos and images, both surgeons and AI developers’ expertise are required. The particular challenges of creating datasets in the surgery field are, collecting the required surgical operation videos which are mostly not in the public domain, selecting the relevant surgery sequences and the most difficult issue is the creation of dataset which represent the full surgical domain spectrum. In my experience the involvement of an expert and experienced surgeon is crucial, defining the required unique cases and the domain diversity (including unique anatomic, techniques, surgical tools and laparoscopy devices). Although there are several algorithms for handling imbalanced dataset, the challenge is still significant and requires the expert surgeon advise.
When using supervised /partially supervised learning methods the annotation phase requires special care. On one hand defining the annotation rules, accuracy verification method, unbiased data handling, and consistency, on the other hand defining the annotators required skills. The annotation phase is expert time consuming and therefore it is important to have a well-defined methodology and close supervision (involving the expert and the AI developer). As a best practice, the annotation should be done by several annotators (expert surgeons) to prevent biased annotations, and requires a good mechanism for dissolving disagreements. The annotation process can also provide an estimation for prediction of complexity (according to the annotators disagreement).
Although the training and testing phases are mostly done by the AI developers, following my experience, in some cases the involvement of the expert surgeon can be helpful during the AI model testing (e.g. as part of model prediction failure investigation). Additional aspect which requires expert surgeon advice, is the set of rules for handling the algorithm prediction accuracy and confidence (e.g. priorities the prediction errors such as False positive or False negative, defining thresholds, etc).
In conclusion, although the time of the expert surgeon is costly, the collaboration between the expert surgeons and AI developers is crucial for the success of AI creation in surgery and should be done throughout all the development process, including specification, design, implementation and testing.
This is one of the most important missions of SurgeonAI in the road to bringing the power of AI to laparoscopic surgery.
Gil Alter – VP R&D – SurgeonAI
For more than two decades, at the forefront of steering software development for state-of-the-art systems, demanding high performance and reliability. Extensive expertise in various domains of software development, encompassing the creation of 3D graphics engines, real-time simulations, computer vision, and AI/ML algorithms. For more than a decade served as a Team leader of infrastructure software projects and leading an AI/ML R&D researchers’ group. As Technical Team Leader, overseen the entire life cycle of state-of-the-art full mission simulator projects. Entrepreneur of a startup company specializing in the automotive realm.