Deep Learning based warning systems – detect bleeding during FESS operations

Problem description:

Bleeding during a FESS operation has a negative impact on the quality of the endoscopic image. Our own examinations could prove, that – depending on the level of experience of the surgeon – in up to 40% of cases the vision was marginally or critically limited by bleeding. This results in an increased risk of surgery due to poor orientation in Situs.

Material and Methods:

In this work, Deep Learning for image recognition and time series analysis is used to develop a warning system which evaluates Video Material of the endoscope during FESS operations. In this way, warning messages are displayed in real-time into the OP cockpit to optimize the operational workflow. Training- and Validation sets have a size of 10.000 and 2000 Images distributed over 1200 FESS operations.

Results:

Using deep learning methods, strong blood flow is detected to alert surgeons to impaired viewing conditions. With 22 different recognized classes, the validation accuracy to correctly recognize bleeding with less than 10% deviation is over 85%.

Conclusion:

The presented assistant system facilitates decisions to adapt the surgical strategy by automatically indicating the degree of bleeding. This allows the surgeon to take reliable measures, like flushing, medical or physical hemostasis. In addition, the system provides the basis for prediction of bleeding events which eventually lead to improved surgical procedures.

Dipl. Geophys. Thomas Grau

Kopfzentrum Gruppe,

Münzgasse 2, 04107

Leipzig

Email: [e-mail]

Publication History

Publication Date:
23 April 2019 (online)