Work

Классификация двумерных изображений хромато-масс-спектрограмм летучих метаболитов мочи с помощью свёрточных нейронных сетей для ранней диагностики онкологических заболеваний

Year: 2024
Type: article
Abstract: Исследована возможность классификации двумерных изображений хромато-масс-спектрограмм летучих метаболитов мочи на предмет наличия онкологических заболеваний с использованием свёрточной нейронной сети. Для этого использован собственный банк данных изображений хромато-масс-спектрограмм, содержащий 108 образцов, полученных от здоровых пациентов и пациентов с клинически подтверждёнными формами рака желудка, рака пищевода, рака лёгкого и рака кишечника. Исходя из особенностей изображений, обоснована архитектура свёрточной нейронной сети для их классификации. С использованием перекрёстной проверки по k блокам получены оценки точности классификации изображений хромато-масс-спектрограмм на предмет наличия или отсутствия онкологического заболевания. Introduction. Introduction. Oncological diseases rank among the leading causes of death globally. The success of treatment significantly hinges on the early detection of tumors. Therefore, a critical objective is the development of cost-effective and less labor-intensive methods for early cancer diagnosis. Different tumor types release distinct substances, which may be detectable in various biological fluids. Utilizing tumor markers presents a potential solution for early cancer diagnosis. Notably, urine, being readily accessible, serves as a valuable source of information on human metabolism, with its collection and analysis being a routine medical procedure. Thus, the identification of specific substances in urine related to particular types of cancer emerges as a pertinent scientific challenge. In this context, gas chromatography-mass spectrometry stands out as a method for quantitatively assessing individual components in urine. The goal of this study was to devise a method for classifying images of gas chromatography-mass spectrometry depicting volatile urine metabolites to detect the presence of oncology. To achieve this, we addressed the following objectives: 1) articulating the scientific foundation for a convolutional neural architecture designed to classify two-dimensional images of gas chromatography-mass spectrometry showcasing volatile urine metabolites; 2) implementing the neural network and its learning algorithm through contemporary deep learning frameworks; 3) evaluating the efficacy of the chosen neural network model in classifying images derived from gas chromatography-mass spectrometry. Findings. Leveraging the features of spectrometry images depicting volatile urine metabolites, we justified the application of a convolutional neural network architecture for their classification in detecting cancer. Employing cross-validation across k blocks, we derived accuracy estimates for classifying images based on the presence or absence of cancer. The neural network underwent training on its proprietary database, comprising 108 gas chromatography-mass spectrogram images, including 32 from healthy subjects, 27 from subjects with stomach cancer, 29 from subjects with lung cancer, 12 from subjects with intestinal cancer, and 8 from subjects with esophageal cancer. The binary image classification accuracy reached 72%, and it's clear that it will further improve with an expanded training set. The neural network completed 40 epochs of training, after which retraining occurred. The practical significance of this work is in the detailed exploration of the results from intermediate convolutional layers of a neural network trained to classify spectrometry images of volatile urine metabolites for cancer detection. This exploration will facilitate the identification of concentration levels of specific metabolites distinctive to each cancer type. Subsequently, these metabolites can be pinpointed using other more accessible analytical methods.
Language: Russian
Open Access status: diamond
OpenAlex ID: W4407290751
FWCI:
Cites: 21
Cited by:
Related to: 10