Identifying brain tumors by differential mobility spectrometry analysis of diathermy smoke

Restricted access

OBJECTIVE

There is a need for real-time, intraoperative tissue identification technology in neurosurgery. Several solutions are under development for that purpose, but their adaptability for standard clinical use has been hindered by high cost and impracticality issues. The authors tested and preliminarily validated a method for brain tumor identification that is based on the analysis of diathermy smoke using differential mobility spectrometry (DMS).

METHODS

A DMS connected to a special smoke sampling system was used to discriminate brain tumors and control samples ex vivo in samples from 28 patients who had undergone neurosurgical operations. They included meningiomas (WHO grade I), pilocytic astrocytomas (grade I), other low-grade gliomas (grade II), glioblastomas (grade IV), CNS metastases, and hemorrhagic or traumatically damaged brain tissue as control samples. Original samples were cut into 694 smaller specimens in total.

RESULTS

An overall classification accuracy (CA) of 50% (vs 14% by chance) was achieved in 7-class classification. The CA improved significantly (up to 83%) when the samples originally preserved in Tissue-Tek conservation medium were excluded from the analysis. The CA further improved when fewer classes were used. The highest binary classification accuracy, 94%, was obtained in low-grade glioma (grade II) versus control.

CONCLUSIONS

The authors’ results show that surgical smoke from various brain tumors has distinct DMS profiles and the DMS analyzer connected to a special sampling system can differentiate between tumorous and nontumorous tissue and also between different tumor types ex vivo.

ABBREVIATIONS CA = classification accuracy; DMS = differential mobility spectrometry; GBM = glioblastoma; LDA = linear discriminant analysis; LGG = low-grade glioma; LOOCV = leave-one-out cross-validation; OCT = optical coherence tomography; REIMS = rapid evaporate ionization mass spectrometry; 10-f-CV = 10-fold cross-validation.

Downloadable materials

  • Supplemental Figs. 1 and 2 (PDF 2.30 KB)

Article Information

Correspondence Ilkka Haapala: Tampere University Hospital, Tampere, Finland. ilkka.haapala@fimnet.fi.

INCLUDE WHEN CITING Published online June 14, 2019; DOI: 10.3171/2019.3.JNS19274.

Disclosures Mr. Karjalainen, Mr. Kontunen, Dr. Oksala, and Dr. Roine: direct stock ownership in Olfactomics Ltd. Dr. Roine: employee of Olfactomics Ltd.

© AANS, except where prohibited by US copyright law.

Headings

Figures

  • View in gallery

    Illustration of the measurement system consisting of a tissue sampling unit (I), sample conditioning unit (II), DMS sensor (III), and computational data analysis (IV). The tissue sampling unit further includes a diathermy power unit (A), robot stage (B), sample well plate (C), surgical smoke evacuator (D), filter and dilution system (E), and ENVI-AMC differential mobility spectrometer (F). The numbered values depict the input flow rate to each device. The LDA is a sketch of the operating principle, where 2 classes, + and o, are projected onto a 2D plane.

  • View in gallery

    Example spectral images from a single control tissue sample (left) and a GBM (right). The areas of the dispersion plot that are in the 1- to 2-V compensation voltage range are the most prone to exhibit the differences in the DMS spectra between the tissues. In this instance, the differential mobility peaks near the 2-V compensation voltage differ in shape and size. VC = compensation voltage; VRF = peak-to-peak amplitude of the radiofrequency waveform voltage.

  • View in gallery

    The principle of LDA classification. The LDA classification model aims to find a linear combination that best separates two or more classes based on given features. In this study, the features were the pixels in the DMS spectra, but for visualization purposes, the data for GBM and LGG are presented in 2 dimensions as Sammon projections of the DMS spectrum data.

TrendMD

Metrics

Metrics

All Time Past Year Past 30 Days
Abstract Views 914 914 84
Full Text Views 110 110 15
PDF Downloads 48 48 8
EPUB Downloads 0 0 0

PubMed

Google Scholar