Intraoperative online image-guided biopsie on the basis of a Deep Learning algorithm to the automatic detection of head and neck carcinoma by means of real time Nah-Infrarot ICG fluorescence endoscopy

Introduction:

Golden standard in the diagnostics and surgery of head neck carcinomas is the investigation with white light. This causes an insufficient tumor-recognition rate with small tumors and above all with big tumors a dissatisfactory characterization of tumor margins. Results are a high recurrence rate of 10 – 30% and a raised mortality conditioned through this. Digital image recognition systems picture a better tumor and tumor margin recognition.

Material and methods:

Within the scope of regular endoscopies of head neck tumors image-guided biopsies with an interactive developed live software by means of a Deep Learning algorithm under real time Near-Infrared Fluorescence Endoscopy were cunducted. The software were coached offline with the results of the histopathologic investigation of the biopsies in annotated pictures.

Results:

The accuracy in the training sphere of the algorithm with the help of histopathologic annotated picture data of 22 patients with the "leave-one-out" method amounted at best to 0.78. Erroneously annotated picture artefacts pose currently still the biggest problem there. Hence, a picture-steered biopsy can occur currently only with correspondent clinical appraisal.

Conclusions:

Intraoperativ arose up to now still no improvement in the recognition of the tumors ore its margin to the healthy mucosa in comparison to the clinical appraisal by an experienced surgeon. The standardised data of 22 patients are not with wide enough. Whether the steadily increasing progress can be utilised in the digital picture processing for the treatment by head neck tumors, must be cleared in substantially bigger, multicentric studies with the help of substantially bigger data amounts.

Publikationsverlauf

Publikationsdatum:
23. April 2019 (online)