Classification and Analysis of Differential Mobility Spectrometry Measurements

Abstract

There are many fields, where chemical analysis and detection volatile organic compounds (VOCs) are needed. Examples are warfare agents monitoring in critical infrastructures, food quality control, and rapid diagnosis of various diseases. The field of chemical detection has evolved, and various approaches were developed for such purposes. One of these approaches is differential mobility spectrometry. The working principle of DMS is based on the behavior of charged molecules in gaseous medium under the effect of weak and strong electric fields. The result of a measurement is a matrix, also called a dispersion plot, with rows that contain measurements over a certain range of separation voltages, and columns that contain measurements over a certain range of so-called compensation voltages. However, the interpretation of DMS measurements is challenging. One of the main challenges is that the DMS measurement does not provide direct information what VOC was measured. Additionally, obtaining information on measured substance by visual inspection is cumbersome or even unfeasible. This thesis aims to overcome these challenges by employing machine learning algorithms for automated analysis. Ongoing research at Tampere University Hospital explores potential of using DMS for brain cancer detection and analysis. DMS might be suitable for analyzing of volatile organic com- pounds emitted by tissues. One application area is surgical treatment of solid cancers. Here, the surgeon must remove malignant cells completely to achieve negative margin, meaning no cancer cells are left at the edges of the excised tissue. A positive margin, in contrast, indicates that cancerous cells remain, increasing the risk of re- currence. However, achieving a negative margin is not always possible because it is difficult to discriminate between healthy and malignant tissue. As a result, patients are often subject to a reoperation, which affects their wellbeing and causes additional costs to health service. Thus, there is a need to develop a method for automatic intraoperative detection of tissue types. Different types of tissues are characterized by different biomolecular content that include proteins, fatty acids, and metabolic products. Cancerous tissues differ from surrounding cells by specific metabolic products and other biomarkers, which can be exploited for automatic tissue analysis. This idea is currently being implemented through the development of an electric knife coupled with DMS. During the incision, the electric knife produces surgical smoke, which is fed into the DMS for rapid analysis. One crucial part of the system, to be addressed in this dissertation, is algorithms for reliable classification and analysis of the tissues being incised.

The aim of this thesis is to develop algorithms and preprocessing methods for analysis and classification through four publications. The work was divided into two parts: the first part focused on the analysis of collected data from measured chemicals, and the second part focused on the processing samples from patients. The results from the first part demonstrated the advantages of approaching the dispersion plots as sequential phenomenon. The first publication in this part discusses the possibility of applying clustering algorithms for isolating the signal in dispersion plots. The results showed that this strategy has potential but needs further investigation. Next, more advanced models for time-series data, such as attention and self-attention mechanisms, and transformers can be tested. The second publication proposes to interpret dispersion plots as multidimensional sequential data and, hence, uses time series analysis algorithms for the classification of dispersion plots. It is shown that this method yields accuracy as high as previously used state-of-the-art algorithms. The papers in the second part show that it is possible to discriminate between isocytrate dehydrogynase mutated cancer and so-called wild type with good accuracy. The publications also confirmed that the use of linear discriminant analysis algorithm is advantageous for this type of tasks. The findings of this thesis offer new perspectives on dispersion plots and contribute to advancing research in the field by introducing novel ideas and methodologies.