Titolo della tesi: AI for safer surgery: generalizing to multicentric laparoscopic cholecystectomy and surgical oncology
Modern surgery is a highly effective but error prone sociotechnical process. Surgical data science is a novel multidisciplinary research field aiming to improve surgical care by modelling surgical data and developing digital tools to assist healthcare providers. Minimally invasive surgical procedures are natively guided by endoscopic images that could be analyzed by deep learning models specifically trained to assist surgeons.
The present doctoral thesis presents a series of clinical, data and computer science studies to prevent bile duct injuries in laparoscopic cholecystectomy (LC). The main contributions towards achieving this goal can be broken down in:
• Assessing LC operative difficulty
• Validating deep learning models for postoperative video documentation of surgical procedures on large multicentric dataset
The difficulty assessment of LCs led to publications on the available score to predict preoperative and intraoperative difficulty and what is the definition of difficult cholecystectomy. In particular, it was proposed and demonstrated that at the moment there is not a common definition of difficult cholecystectomy and only few scores are being validated. Moreover, we test the available intraoperative score on a large video dataset showing poor performance
Finally, deep learning models analyzing LC videos to segment hepatocystic anatomy and to selectively video document LC safety (EndoDigest) were developed and tested on a large multicentric video dataset. We try also to export our model for the LC to oncological surgery, specifically in right hemicolectomy but we found some issue particularly in the standardization of the procedure.
In conclusion, the works presented in this thesis lay the foundation to use surgical data and computer science analytics to make cholecystectomy, one of the most performed abdominal surgical procedures, safer.