Learning-based Surgical Workflow Detection from Intra-Operative Signals

Thesis by: Ergün Kayis
Advisor: Prof. Nassir Navab
Supervision by: Ralf Stauder, Aslı Okur, and Loïc Peter
Timeframe: six months

Abstract

Systematic knowledge of surgical activities is very helpful for various scenarios, as this can aid in an automatic generation of medical report drafts, it can be utilized for objective skill assessments or steps following the surgery can be initialized based on the estimated remaining intervention time. Therefore Surgical Process Modelling (SPM) and especially the online recognition of workflow phases is a field of growing interest, as more and more devices and signals become available.

Different approaches and methods have been implemented in the past to recognize workflow phases on recorded surgeries or predict the current phase in surgeries, though so far none of these methods could be applied to a real scenario, as the required data was never available on site. With the availability of RFID- and other signals in actual surgeries, it is now possible to attempt the detection of workflow phases in live interventions. This will enable many other systems to build upon the online availability of workflow data and allow for further development of context-sensitive systems.

The goal of this project will be to apply methods from Machine Learning (ML) to medical data sets in order to deduct the current workflow phase. These data sets were recorded by our medical partners during actual laparoscopic cholecystectomies and will contain binary values (like the usage vector of all possible surgical instruments) as well as analog measurements (e.g. intra-abdominal pressure). By learning from labeled data, methods like Random Forests or Hidden Markov Models should be able to detect which of the known phases is the most probable, given the data at hand.