Artikel
Novel rapid intraoperative tumour detection for resection and biopsy control by a residual convolutional neural network using stimulated raman scattering microscopy
Neuartige schnelle intraoperative Tumordetektion zur Resektions- und Biopsiekontrolle durch ein residuelles neuronales Netzwerk unter Nutzung stimulierter Raman-Streuung-Mikroskopie
Suche in Medline nach
Autoren
-
David Reinecke
- Universitätsklinikum Köln, Klinik für Allgemeine Neurochirurgie, Köln, Deutschland
-
Niklas von Spreckelsen
- Universitätsklinikum Köln, Klinik für Allgemeine Neurochirurgie, Köln, Deutschland
-
Christian Mawrin
- Universitätsklinikum Magdeburg, Institut für Neuropathologie, Magdeburg, Deutschland
-
Adrian Ion-Margineanu
- Invenio Imaging Systems Inc., Santa Clara, CA, Vereinigte Staaten
-
Gina Fürtjes
- Universitätsklinikum Köln, Klinik für Allgemeine Neurochirurgie, Köln, Deutschland
-
Stephanie T. Jünger
- Universitätsklinikum Köln, Klinik für Stereotaxie und funktionelle Neurochirurgie, Köln, Deutschland
-
Florian Khalid
- Invenio Imaging Systems Inc., Santa Clara, CA, Vereinigte Staaten
-
Christian Freudiger
- Invenio Imaging Systems Inc., Santa Clara, CA, Vereinigte Staaten
-
Marco Timmer
- Universitätsklinikum Köln, Klinik für Allgemeine Neurochirurgie, Köln, Deutschland
-
Maximilian I. Ruge
- Universitätsklinikum Köln, Klinik für Stereotaxie und funktionelle Neurochirurgie, Köln, Deutschland
-
Roland Goldbrunner
- Universitätsklinikum Köln, Klinik für Allgemeine Neurochirurgie, Köln, Deutschland
-
Volker Neuschmelting
- Universitätsklinikum Köln, Klinik für Allgemeine Neurochirurgie, Köln, Deutschland
| Veröffentlicht: | 25. Mai 2022 |
|---|
Gliederung
Text
Objective: The determination of the presence of tumor in biopsies and the decision-making during resections is often dependent on intraoperative histological tissue sample assessment. Recently, stimulated Raman scattering microscopy has been introduced to rapidly generate virtual hematoxylin-and-eosin-stained-like images (SRH) for intraoperative brain tumor diagnosis. For the prediction of tumor presence, we aimed to develop a new deep residual convolutional neural network (CNN) and test its reliability in an automated clinical workflow.
Methods: The CNN was established and trained in ResNetV50 to predict three classes for each SRH: (1) tumor, (2) non-tumor and (3) low quality. In a monocentric prospective clinical study with 94 consecutive patients who underwent biopsy or brain tumor resection the CNN was applied on SRH images obtained in three 4 mm2 random areas within the squeezed tissue samples. SRH images were blindly independently reviewed by a neuropathologist serving as ground truth. Internal consistency was calculated using Cronbach's alpha (Cα) and interrater reliability between CNN and neuropathologist by Cohens Kappa (κ).
Results: In total, 402 SRH images deriving from 132 tissue samples were analyzed. The automated workflow took in mean 240 seconds per case in total, the CNN could successfully be applied in all cases. In the majority (90.0%) of SRH images the CNN could correctly classify tumor (n=303) and differentiate from normal brain tissue (n=49) in accordance to the neuropathologist at a mean probability value level of 78.2 ±31.6 %. The reliability between CNN and neuropathologist showed a substantial interrater agreement in 90% (κ=0.647). There was an excellent consistency found among the random areas within one tissue sample with 90.9% (Cα=0.93) accuracy. The pathological diagnoses represented the full spectrum of neurooncological surgery.
Conclusion: The newly developed CNN coupled to SRH in an automated workflow can reliably detect the microscopic presence of tumor tissue in resection and biopsy samples within 4 minutes in total and may pave a promising way for an alternative rapid intraoperative decision-making tool.
Figure 1 [Fig. 1]
