[MA] Architecture Design for Extendable Productive ML Systems on Video Data
Motivation
Since the late 1960s, a new and innovative procedure called ERCP (Endoscopic Retrograde
Cholangiopancreatography) has been introduced in the medical field and has become part
of the toolkit of doctors ever since. This procedure, as deduced by its name, is performed by
putting a flexible tube called “endoscope” through the mouth of the patients, into the stomach
and a part of the small intestine called “duodenum”. By attaching a camera in the en-
doscope, it allows doctors to see small tubes inside the body called pancreatic and bile ducts
that carry digestive juices to the intestines. These ducts are particularly interesting if doctors
suspect a disease of the pancreas or liver, such as stones, lesions or tumors.
ERCP has become a de-facto standard procedure, as it proves to be less invasive compared to
alternatives like open or laparoscopic surgery and highly successful. But even with a high suc-
cess rate, ERCP is not a walk in the park. Under certain factors like rearrangement of the up-
per gastrointestinal anatomy of the patient , the access to the duct might be limited or
even prevented – this makes the ERCP quite complex and non-trivial, especially for non-senior
doctors. In these cases, the physicians usually fail to enter or even locate the duct, and thus the
procedure will fail.
Thankfully, we are living in the middle of the so-called technology “boom”. Huge ad-
vancements have been made in the way we capture, store, process, and analyze data. With high-
quality and adaptable video-cameras, data lakes that store colossal amounts of data, distributed
systems that can process in parallel and real-time, and techniques like Artificial Intelligence and
Machine Learning that provide additional “brainpower” with little effort, there is a wide pool of
potential use cases that can aid humanity strive forward in many areas, and the ERCP proce-
dure is one potential candidate.
Lately, in compliance with data protection laws, the ERCP procedure is being recorded and the
resulting videos are being stored for future use. This brings forward a great potential use-case.
With a huge dataset pool of ERCP videos, the mentioned technological advancements can po-
tentially provide aid for physicians in future non-trivial ERCP procedures - a technical proposal
to solve this problem will be the main focus of the thesis.
Goal
As discussed in the introduction, complicated ERCP procedures bring up real issues to phy-
cisians that need to be addressed and if possible fixed with the help of technology. Some con-
crete pain points include:
Under a rearranged upper gastrointestinal anatomy of the patient, locating the ducts is
non-trivial.
After locating, the access to the duct might be limited or even prevented.
Upon entry to the ducts, small but existing risks for pancreatitis, cholangitis, hemorrhage,
and duodenal perforation still exist .
With these pain points in mind, the goal of this thesis is to design a system, a live-
supporting tool, powered by trained ML/AI models, that recognizes pain-points
during the procedure and helps guide the physicians in any difficulties during ERCP.
Besides that, it will enable technicians to evaluate and integrate technology itera-
tively.