Endoscopic submucosal dissection (ESD) enables minimally invasive resection of large lesions and thus offers an organ-preserving alternative to surgical resection. However, ESD is a technically demanding procedure with considerable risks: The unintentional cutting of blood vessels prior to their coagulation can lead to severe bleeding, which not only impairs the surgeon's vision but can also endanger the patient.

Our SmartESD deep learning system supports the surgeon by detecting and highlighting blood vessels in the endoscopic video image in real time. At the same time, the position of the dissection knife is continuously tracked. If the knife comes too close to critical structures such as blood vessels, the system proactively warns the surgeon.

SmartESD thus acts as a vigilant surgical assistant that increases the safety of the procedure and can contribute to lower complication rates - particularly valuable in complex anatomical situations or during longer procedures where concentration can wane.

Endoscopic submucosal dissection typically goes through several defined phases: Diagnosis, marking, injection, dissection and hemostasis. The sequence, duration and frequency of these phases are characteristic of the course and quality of a procedure. While an ESD with few complications is characterized by few bleeding episodes and continuous dissection progress, frequent changes between dissection and hemostasis indicate technical difficulties or complications.

Our deep learning system automatically recognizes these phases in the endoscopic video and thus creates a precise "fingerprint" of the procedure. This objective documentation enables a wide range of applications: Interventions can be compared and evaluated using quantitative quality metrics - for example, in terms of the total duration of bleeding phases or the efficiency of the dissection. For doctors in training, automatic phase recognition can make individual learning curves visible: With increasing experience, the quality metrics should continuously improve. In addition, particularly challenging ESDs can be identified so that they can be processed for training purposes.