Loading...

Evaluation of the Potential of Deep Learning Methods for Qualitative and Quantitative Analysis of Mass Spectrometry Images

Golpelichi, Fatemeh | 2021

457 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 53606 (03)
  4. University: Sharif University of Technology
  5. Department: Chemistry
  6. Advisor(s): Parastar Shahri, Hadi
  7. Abstract:
  8. In recent years, studying of biological tissues by mass spectrometry imaging (MSI) has been considered due to its selectivity in identifying different compounds in biological tissues, no need for sample preparation, and the possibility of creating the distribution map of these compounds. The complexity of biological tissues due to their heterogeneity, the large volume of data generated, and the effects of competition of other species for ionization in MSI experiments have doubled the importance of using chemometrics to interpret these data. The aim of this work is to quantitatively study Chlordcone as a carcinogenic pesticide and to extract its spatial distribution pattern in mouse liver using mass spectrometry imaging with matrix assisted desorption ionization (MALDI-MSI) deep learning (DL) data analysis methods and for the first time, a convolutional neural network (CNN) has been used to quantitatively analyze MALDI-MSI data. For this purpose, data from 7 standard spots containing 0 to 20 picomoles of Chlordecone and four unknown spots from the mouse liver infected with chlordecone for 1, 5 and 10 days were analyzed. First, the data were compressed by binning method with three bin sizes of 1, 0.5 and 0.25. Standard spot data were unfolded in matrix form and stacked. To solve the lack of sufficient labeled data for CNN model training, multivare curve resolution-alternating least squares (MCR-ALS) method has been implemented to extract the pure Chlordcone spectrum to convert it to concentration labels for stabdards data pixels. The designed CNN model was trained using 918, 1260, and 1140 labeled pixels for 1, 5, and 10 days data, respectively and then models was evaluated using 102, 140, and 127 labeled pixels as test data, respectively. Prediction R2 for all three dataset, ranged from 0.91 to 0.97 in three different bin sizes. Also, prediction and calibration R2 was compared with SVM and PLS models which CNN was superior in all cases. Trained CNN models were used to predict the amount of Chlordecone per pixel, per spot and to visualize its spatial distribution in mouse liver tissues. The 0.25 bin size has produced more accurate results. The prediction results of CNN model were compared with univariate MALDI-MSI and GC-MS methods and the values obtained from CNN were higher than MALDI-MSI and lower than GC-MS. The results of this study demonstrate the power of deep learning methods for modeling MALDI-MSI data for quantitative analysis, minimizing the effects of tissue heterogeneity on standard and unknown data, and predicting unknown values of Chlordecone in mouse liver
  9. Keywords:
  10. Machine Learning ; Deep Learning ; Chemometrics Method ; Chlordecone ; Mass Spectrometry Imaging (MSI) ; Convolutional Neural Network

 Digital Object List

 Bookmark

No TOC