Deep Learning-Based Synthesis of Surgical Hyperspectral Images

Postoperative death within 30 days after surgical intervention is the third largest contributor to mortality globally. Causes of postoperative mortality are manifold but also comprise challenging perception and the inability to estimate physiological tissue parameters during interventions. To capture data emanating from underlying physiological tissue properties, hyperspectral imaging (HSI) together with machine learning-based analyses has been proposed as a solution in recent literature. However, HSI data in the clinical setting is sparse, as its acquisition is crucially limited by a small number of approved devices and the need for clinical trials. Therefore, the present work investigates common deep learning frameworks for HSI and proposes a two-step image generation pipeline to synthesize hyperspectral tissue images. To validate the image generation pipeline, spectral correctness and textural realism were assessed both qualitatively and quantitatively. Results of the textural Kernel Inception Distance (KID) exhibited state of the art (SOTA) performance for both paired and random generated HSI patches. Furthermore, the feasibility of using the synthetic, unlabelled data for an image segmentation task was tested and found to not lead to improvement. From the conducted experiments it can be concluded that RGB image synthesis can be adapted to the HSI domain, while synthetic additional data has to be tailored for individual tasks.

Level thesis:degree_level

master

Grantor dc:publisher

Universität Heidelberg

Year

2021

Author dc:creator

• Hübner, Marco

Contributors dc:contributor

• Rother, Carsten

Repository record source_url

http://www.ub.uni-heidelberg.de/archiv/31421

OAI identifier oai:identifier

oai:archiv.ub.uni-heidelberg.de:31421